Dev.to AI 🤖 Ai 👁 0 📖 35 min read

Inkling 2 Mini vs Qwen3.8-27B and Arete-27B on agent tasks

An agent that sorts a support inbox or places orders has to call the right tool at every step, and a single call the tool refuses can leave it retrying until its step limit runs out. Inkling 2 Mini, an open model that Th

Inkling 2 Mini vs Qwen3.8-27B and Arete-27B on agent tasks

An agent that sorts a support inbox or places orders has to call the right tool at every step, and a single call the tool refuses can leave it retrying until its step limit runs out. Inkling 2 Mini, an open model that Thinking Machines released on 20 November 2026, is a mixture of experts that uses 3.4 billion of its 26.4 billion parameters for each token, so it could run an agent faster than a 27-billion-parameter model, if it calls tools as reliably.

We put Inkling 2 Mini through the five agent tasks we use for our own models, beside Qwen3.8-27B and Arete-27B, two 27-billion-parameter models we serve. Each model made 54 runs at its own model card's settings: nine task variants (four tasks in Polish and English, one in Polish only), each three times with a short and three times with a guided prompt. What we found:

  • with a step-by-step procedure, Inkling 2 Mini passed all 27 runs and the two 27-billion-parameter models 21 each, stalled by a strict tool that takes no arguments until it ignored the argument they invented;
  • with a short prompt, it passed 20 of 27 runs against 26 for Qwen3.8-27B and 27 for Arete-27B, mostly by asking the user before accepting a higher price: right for an agent that must not spend on its own, slower for one that may;
  • on one RTX 5090 it wrote 161 tokens a second against 84 and 68 (4-bit builds, no draft model) and finished an agent task in 14.4 seconds at the median against 18.2, so one card serves more users sooner;
  • on a third of its requests the time to first token was 1 to 4 seconds, against under a second for the other two: a pause that a person waiting on a chat agent notices and a batch job hardly does;
  • its published image for vLLM, the open-source server it runs on, lacks the code for H100 and H200 cards, so teams with those cards cannot run it on that image yet.

The runs show how each model behaves on such work, not a benchmark score: five small tasks, three runs each, on two graphics cards. A benchmark would need more tasks and runs, more kinds of card and prompts tuned for each model.

Developers choosing an open model for tool calling will see where each one fails and what prevents it, teams running models on their own card the speeds, and both how to pick models for further testing; every model's work is quoted verbatim at the end. Our guide to Bielik 11B v3.1 through an API shows how to write such guided prompts.

Three of the strongest small open models

On the Artificial Analysis Intelligence Index (version 5.0.1), Inkling 2 Mini and Qwen3.8-27B both score 20, the highest of any open-weights model under about 40 billion parameters. The top models score 62 to 65 and the best open-weights model 46, so these are the strongest models that fit one graphics card. Arete-27B, Nous Research's fine-tune of Qwen3.8-27B, is not on the index. The three differ more in the market than in the score:

Inkling 2 Mini Qwen3.8-27B Arete-27B
Made by, released Thinking Machines, November 2026 Qwen, August 2026 Nous Research, November 2026
Parameters, in all / per token 26.4 billion / 3.4 billion 27 billion / 27 billion 27 billion / 27 billion
Intelligence Index 20 20 not listed
Downloads, last 30 days 23,500 6.3 million 17,700
Paid hosts on OpenRouter 2 23 none
Price per million tokens, input / output $0.075 to $0.08 / $0.30 to $0.35 from $0.035 / from $1.05 none listed

A team that wants a hosted API has 23 hosts to choose from for Qwen3.8-27B and two, cheaper per output token, for Inkling 2 Mini, which on a team's own card runs on a preview build of vLLM for now; Arete-27B its users run themselves. Neither the score nor the market shows how a model copes when a tool refuses a call, which is what the tasks test.

Five agent tasks in Polish and English, run as an app would

Each task gives a model a few tools, the data of a made-up company behind them (its two travel policies and staff records, for example) and a request, and checks the result against answers known in advance:

Task What the agent does The trap
Travel policy answers a question from a document store and the HR system an older policy with other numbers
Support inbox reads six emails, files tickets for customers' complaints a newsletter, spam, a sender who is not a customer
Meeting and room finds the earliest hour three people are free, books a room a 30-minute gap, a room already taken
Order places an order with a substitute for a short item the shop refuses the first attempt over a price change
Overdue invoices finds a customer's overdue invoices, saves a reminder (Polish only) a similar company name, a partly paid invoice

The tasks test following a procedure and recovering from errors in three to fourteen tool calls, not knowledge, long documents or code. Each runs with a short system prompt, a role and "don't guess data, check it with the tools", and a guided one that adds definitions, a numbered procedure with one tool call per step, and rules such as no text between tool calls. A loop like a customer's app runs the tool calls and hands back the results, for at most 16 requests.

Each model ran at its model card's settings, reasoning before each answer and sending its earlier reasoning back at each step, as Qwen's model card advises, on one RTX 5090, a 32 GB graphics card. Qwen3.8-27B and Arete-27B also ran on an H200 NVL, the data-centre card on which Blixild AI serves them; Inkling 2 Mini's only published vLLM image lacks the code for that card, an open issue at vLLM. Every time was measured on the machine itself.

Inkling 2 Mini runs in the precision it was trained in, mostly 4-bit, so its build loses nothing to conversion. Qwen3.8-27B and Arete-27B were released in 16 bits and ran in quantized builds, their weights stored in fewer bits to save memory and time, so their results hold for these builds, not for the models at full precision. A draft model, which guesses several tokens for each pass of the full model (speculative decoding), ran only where the table says:

Model Weights in the build, and who made it Cache Graphics card Draft model
Inkling 2 Mini NVFP4, 4-bit floating point for the experts, 8-bit (MXFP8) for the last four layers, 16-bit attention, as trained; Thinking Machines' own release 16-bit RTX 5090 none (it did not start)
Arete-27B W4A16, 4-bit integers in most layers with 16-bit arithmetic; a third party's build (supz) 8-bit RTX 5090, H200 NVL none
Qwen3.8-27B NVFP4, 4-bit floating point, with 8-bit (FP8) attention; NVIDIA's build 8-bit RTX 5090 none (built-in, for one speed figure)
Qwen3.8-27B FP8, 8-bit floating point; Qwen's own build 8-bit H200 NVL built-in (MTP)

Within that scope, the clearest difference showed with the guided prompts.

With a guided prompt, Inkling 2 Mini passed every run

Agent-task runs that passed every check
Measured. Of 27 runs per model with each kind of prompt, five tasks in Polish and English with three runs each, how many passed every check. Inkling 2 Mini ran on an RTX 5090, the other two on an H200 NVL.

With the procedure written out, all three models got the policy, the meeting, the order and the reminder right in every run, and Inkling 2 Mini the inbox too: three tickets, high priority for the customer whose production had stopped, and the non-customer left for a person.

Qwen3.8-27B and Arete-27B failed the inbox in all twelve runs. Its first tool, which lists the emails, takes no arguments; both models gave it one, such as {"void": "true"}, the tool refused it, and they repeated the call, mostly unchanged, until the run ended without an answer. With the short prompt, both called it correctly every time.

The cause is the strict tool: when it ignored arguments it does not take, as many APIs ignore fields they do not know, both models still invented one but sorted the inbox right in all six runs each. A tool that ignores unknown fields is a cheap guard against such a stall.

One tool in one task shows a way an agent can stall, not how often. On the RTX 5090, in 4-bit builds, Qwen3.8-27B passed 48 of 54 runs and Arete-27B 46, against 47 and 48 on the H200, failing the same inbox, so the card and the build did not change the finding. The short prompts turned the ranking around.

With a short prompt, Inkling 2 Mini asks before it accepts a higher price

With only a role and one rule, Qwen3.8-27B passed 26 of 27 runs and Arete-27B all 27. Inkling 2 Mini passed 20; five of its seven misses came from the order task.

There the shop refused the first order because the substitute's price had risen from 125.00 to 129.00, and asked for a confirmation flag. Qwen3.8-27B and Arete-27B added it, placed the order for 894.00 and reported the substitute, in all twelve runs. Inkling 2 Mini did so once; in the other five runs it asked: "Do you want me to confirm and place the order with the updated B-210 price (129.00), or do you want something else?"

Our check counted that as a failure, as the task expected the order placed, though the request did not say whether the agent may accept a higher price. Given the guided procedure, which says to fix what an error asks and call again, Inkling 2 Mini placed it in all six runs: an agent allowed to accept a higher price needs a prompt that says so.

Its other two misses, and Qwen's one, were in the Polish invoice task: the right invoices and total, but a reminder opening "Dzień dobry" where our check expected the contact's name, which only the guided prompt asks for. Inkling's model card warns of lower quality outside English; it made an occasional Polish slip ("faktuarach"), and Qwen and Arete one each ("Uprzejmий", with Cyrillic letters; "Zapiszłem"), so text a customer reads still needs a spelling check or a person's look.

Generation speed on one RTX 5090, and on an H200 NVL

Each model wrote 600 tokens for the same two questions, one in English and one in Polish, at its model card's sampling; the last two columns come from the agent runs:

On one RTX 5090 One request, EN / PL Eight at once, each Eight together Agent task, median First token, longest
Inkling 2 Mini, no draft 161 / 161 tok/s 95 / 93 tok/s 740 / 726 tok/s 14.4 s 4.1 s
Arete-27B, 4-bit, no draft 84 / 84 tok/s 78 / 78 tok/s 606 / 604 tok/s 18.2 s 0.9 s
Qwen3.8-27B, 4-bit, no draft 68 / 68 tok/s 63 / 63 tok/s 491 / 490 tok/s 18.2 s 0.3 s
Qwen3.8-27B, 4-bit, built-in draft 136 / 133 tok/s 130 / 115 tok/s 503 / 452 tok/s not run not run

Without a draft model, Inkling 2 Mini wrote about twice as fast as the two 27-billion-parameter models for one request and a fifth to a half faster with eight at once. With its draft model, Qwen3.8-27B came within a sixth of it for one request, though not with eight.

Inkling's lead in the agent tasks is smaller, because on 121 of its 357 requests the first token took 1 to 4 seconds, mostly 2 to 3, where the other two took at most 0.9 and 0.3. The wait was on the server, while it read the prompt, not in the network; it did not grow with the prompt's length, and its cause is not yet known.

On the H200 NVL, where we serve the two 27-billion-parameter models:

On one H200 NVL One request, EN / PL Eight at once, each Eight together Agent task, median First token, longest
Qwen3.8-27B, built-in draft 189 / 168 tok/s 162 / 146 tok/s 1,070 / 1,044 tok/s 6.9 s 0.3 s
Arete-27B, no draft 91 / 91 tok/s 85 / 84 tok/s 662 / 660 tok/s 15.7 s 0.3 s

There Arete-27B wrote about 8 percent faster than on the RTX 5090, and the 8-bit Qwen3.8-27B with its draft model, as our API serves it, finished these tasks fastest, in 6.9 seconds at the median. No other requests shared the cards, except another test during about six minutes of Inkling's agent runs, which can only have slowed them.

How to pick models for your agent's own tests

Five small tasks cannot choose a model for anyone's agent, and no list of steps fits every agent: how much testing one needs depends on what it can do and what its mistakes would cost, from a draft a person reads to an order it places on its own. What runs like these can offer is a way to pick the models worth testing further:

  1. Keep the models whose model cards and licences fit your use and whose speed on your graphics card or API you can accept.
  2. Try each on a few dozen of your own requests, including ones where a tool fails or the agent should ask first, with several runs each, since a single run can pass by chance.
  3. Read how each one failed: a strict tool, a rule missing from the prompt, or a step the model cannot do. The first two are fixed in your tools or prompt; the last rules the model out.
  4. Take the models that remain into the testing your agent needs, set by its risks: more cases and runs, a person's review where mistakes are costly, and the tests repeated after every change and while it runs.

Such a shortlist holds only for the requests you tried: it says a model is worth the full test, not that it passed one.

Inkling 2 Mini runs in Blixild AI on our RTX 5090 at $0.10 per million input tokens and $0.50 per million output tokens; Qwen3.8-27B and Arete-27B run on our H200 NVL at $0.20 and $0.10 per million input tokens and $1.50 per million output tokens, all in Poland. A new account starts with $1 of credit. We set up and run models like these at a customer's site or in their own cloud account as part of our AI workload hosting.

Every answer, verbatim

For each task, one setting the text discusses: the request, then for each model every tool call (→) with what the tool sent back (←), the data it worked from, and its final answer, all as they came back; the first of three runs, except Qwen's inbox run, the second.

Travel policy question, short prompt, English

The request:

Anna Kowal is going on a business trip to Gdańsk from 2 to 6 December 2026 and will spend 4 nights in a hotel. What is the most she can claim for the hotel in total under the current rules, and who must approve the trip? Give the amount, the approver's name and the document you based it on.

Inkling 2 Mini, 7 of 7 checks

Calls and returns:

→ get_employee({"name":"Anna Kowal"})
← {"name": "Anna Kowal", "grade": "J2", "department": "Sprzedaż / Sales", "team_lead": "Piotr Nowak", "head_of_department": "Ewa Zielińska"}
→ search_documents({"query":"business trip hotel allowance rules"})
← {"results": [{"id": "D-214", "title": "Business travel policy (2026)", "date": "2026-01-01", "start": "This policy replaces the 2024 travel policy from 1 January 2026. Hotel limit per night: ju"}, {"id": "D-101", "title": "Business travel policy (2024)", "date": "2024-02-01", "start": "Hotel: up to 400 PLN per night for all employees. Approval: every business trip is approve"}]}
→ read_document({"id":"D-214"})
← {"id": "D-214", "title": "Business travel policy (2026)", "date": "2026-01-01", "text": "This policy replaces the 2024 travel policy from 1 January 2026. Hotel limit per night: junior grades (J1-J3) 350 PLN; senior grades (S1-S3) 500 PLN. Approval: trips of up to 2 days are approved by the team lead; trips of 3 days or more by the head of department. A trip's length in days counts the departure and return days."}

Answer:

**Total hotel claim: 1,400 PLN** (350 PLN/night × 4 nights)

**Approver: Ewa Zielińska** (Head of Department, Sprzedaż / Sales)

**Basis:** Business travel policy (2026), document D-214 (effective 1 January 2026, replacing the 2024 policy).

How it adds up:
- Anna Kowal is grade **J2** (junior grade, J1–J3) → hotel limit **350 PLN per night**.
- A trip from 2 to 6 December 2026 counts as **5 days** (departure and return days included), which is 3 or more days → approval required from the **head of department**, not the team lead. Anna's head of department is **Ewa Zielińska**. (The 2026 policy puts trips of up to 2 days with the team lead — Piotr Nowak — but that threshold doesn't apply here.)
- 4 nights × 350 PLN = **1,400 PLN**.

Note: the 2024 policy (400 PLN/night) is no longer in force — the 2026 policy governs.

Qwen3.8-27B, 7 of 7 checks

Calls and returns:

→ search_documents({"query": "business trip hotel expense reimbursement limit"})
← {"results": [{"id": "D-214", "title": "Business travel policy (2026)", "date": "2026-01-01", "start": "This policy replaces the 2024 travel policy from 1 January 2026. Hotel limit per night: ju"}, {"id": "D-101", "title": "Business travel policy (2024)", "date": "2024-02-01", "start": "Hotel: up to 400 PLN per night for all employees. Approval: every business trip is approve"}]}
→ search_documents({"query": "business trip approval authorization"})
← {"results": [{"id": "D-214", "title": "Business travel policy (2026)", "date": "2026-01-01", "start": "This policy replaces the 2024 travel policy from 1 January 2026. Hotel limit per night: ju"}, {"id": "D-101", "title": "Business travel policy (2024)", "date": "2024-02-01", "start": "Hotel: up to 400 PLN per night for all employees. Approval: every business trip is approve"}]}
→ read_document({"id": "D-214"})
← {"id": "D-214", "title": "Business travel policy (2026)", "date": "2026-01-01", "text": "This policy replaces the 2024 travel policy from 1 January 2026. Hotel limit per night: junior grades (J1-J3) 350 PLN; senior grades (S1-S3) 500 PLN. Approval: trips of up to 2 days are approved by the team lead; trips of 3 days or more by the head of department. A trip's length in days counts the departure and return days."}
→ get_employee({"name": "Anna Kowal"})
← {"name": "Anna Kowal", "grade": "J2", "department": "Sprzedaż / Sales", "team_lead": "Piotr Nowak", "head_of_department": "Ewa Zielińska"}
→ calculate({"expression": "350 * 4"})
← {"result": 1400}

Answer:

**Answer:**

- **Maximum hotel claim: 1,400 PLN** — 4 nights × 350 PLN per night. Under the current 2026 policy, junior grades (J1–J3) are capped at 350 PLN per night, and Anna Kowal is grade **J2**.
- **Approver: Ewa Zielińska** (Head of Department, Sales). The trip runs 2–6 December 2026, which counts as **5 days** (departure and return days included), i.e. 3 days or more — so approval must come from the head of department, not the team lead (Piotr Nowak).
- **Basis: Document D-214, "Business travel policy (2026)"**, dated 2026-01-01 — it replaced the 2024 policy (D-101) from 1 January 2026, so it applies to this trip.

Arete-27B, 7 of 7 checks

Calls and returns:

→ search_documents({"query": "business trip hotel per diem expenses Gdańsk"})
← {"results": [{"id": "D-214", "title": "Business travel policy (2026)", "date": "2026-01-01", "start": "This policy replaces the 2024 travel policy from 1 January 2026. Hotel limit per night: ju"}, {"id": "D-101", "title": "Business travel policy (2024)", "date": "2024-02-01", "start": "Hotel: up to 400 PLN per night for all employees. Approval: every business trip is approve"}, {"id": "D-187", "title": "Domestic per diem", "date": "2025-06-15", "start": "Per diem for a day of domestic travel: 45 PLN. With breakfast provided, it drops by 25%."}]}
→ search_documents({"query": "business travel approval rules"})
← {"results": [{"id": "D-101", "title": "Business travel policy (2024)", "date": "2024-02-01", "start": "Hotel: up to 400 PLN per night for all employees. Approval: every business trip is approve"}, {"id": "D-214", "title": "Business travel policy (2026)", "date": "2026-01-01", "start": "This policy replaces the 2024 travel policy from 1 January 2026. Hotel limit per night: ju"}, {"id": "D-187", "title": "Domestic per diem", "date": "2025-06-15", "start": "Per diem for a day of domestic travel: 45 PLN. With breakfast provided, it drops by 25%."}]}
→ read_document({"id": "D-214"})
← {"id": "D-214", "title": "Business travel policy (2026)", "date": "2026-01-01", "text": "This policy replaces the 2024 travel policy from 1 January 2026. Hotel limit per night: junior grades (J1-J3) 350 PLN; senior grades (S1-S3) 500 PLN. Approval: trips of up to 2 days are approved by the team lead; trips of 3 days or more by the head of department. A trip's length in days counts the departure and return days."}
→ get_employee({"name": "Anna Kowal"})
← {"name": "Anna Kowal", "grade": "J2", "department": "Sprzedaż / Sales", "team_lead": "Piotr Nowak", "head_of_department": "Ewa Zielińska"}
→ calculate({"expression": "350 * 4"})
← {"result": 1400}

Answer:

**Answer:**

- **Hotel claim:** **1,400 PLN** total — Anna Kowal is grade **J2** (junior, J1–J3), so the limit is **350 PLN per night** × 4 nights = 1,400 PLN.
- **Approver:** **Ewa Zielińska**, Head of Department. The trip runs 2–6 December = **5 days** (departure and return days count), so trips of 3+ days require the head of department, not just the team lead (Piotr Nowak).
- **Document:** **D-214 – "Business travel policy (2026)"** (effective 1 January 2026; replaces the 2024 policy).

Support inbox, guided prompt, English

The request:

Go through the support inbox. For each email that is a complaint or a problem report from a customer in our system, create a ticket: category defect, delivery or billing; priority high only if the customer says their work or production has stopped, normal otherwise. Skip newsletters and spam. If the sender isn't in the customer system, don't create a ticket: list them for me at the end. Then give me a short summary.

Inkling 2 Mini, 7 of 7 checks

Calls and returns:

→ list_inbox({})
← {"emails": [{"id": "E-1", "from": "[email protected]", "subject": "Router stopped working after 2 weeks"}, {"id": "E-2", "from": "[email protected]", "subject": "Tech Weekly #48: this week's news"}, {"id": "E-3", "from": "[email protected]", "subject": "Wrong amount on invoice FV/11/2026"}, {"id": "E-4", "from": "[email protected]", "subject": "You have won an iPhone!"}, {"id": "E-5", "from": "[email protected]", "subject": "Delivery missing - our production has stopped"}, {"id": "E-6", "from": "[email protected]", "subject": "Printer complaint"}]}
→ read_email({"id":"E-1"})
← {"id": "E-1", "from": "[email protected]", "subject": "Router stopped working after 2 weeks", "body": "Hello, the router we bought on 3 November no longer switches on. Please repair or replace it. Jan Nowak, Budmax"}
→ read_email({"id":"E-2"})
← {"id": "E-2", "from": "[email protected]", "subject": "Tech Weekly #48: this week's news", "body": "This week: new processors, a laptop review and 10 security tips. Unsubscribe here."}
→ read_email({"id":"E-3"})
← {"id": "E-3", "from": "[email protected]", "subject": "Wrong amount on invoice FV/11/2026", "body": "Invoice FV/11/2026 says 4,200 PLN but the quote was 3,800 PLN. Please correct it. Ortopedix accounts"}
→ read_email({"id":"E-4"})
← {"id": "E-4", "from": "[email protected]", "subject": "You have won an iPhone!", "body": "Congratulations! Click the link and enter your card details to claim your prize."}
→ read_email({"id":"E-5"})
← {"id": "E-5", "from": "[email protected]", "subject": "Delivery missing - our production has stopped", "body": "The order from 10 November was due yesterday. It hasn't arrived and our production line has stopped. Urgent! Kartonpol"}
→ read_email({"id":"E-6"})
← {"id": "E-6", "from": "[email protected]", "subject": "Printer complaint", "body": "The printer I bought in your online shop prints stripes. I want to return it. Marta Kowalczyk"}
→ find_customer({"email":"[email protected]"})
← {"found": true, "customer_id": "C-201", "name": "Budmax"}
→ find_customer({"email":"[email protected]"})
← {"found": true, "customer_id": "C-305", "name": "Ortopedix"}
→ find_customer({"email":"[email protected]"})
← {"found": true, "customer_id": "C-118", "name": "Kartonpol"}
→ find_customer({"email":"[email protected]"})
← {"found": false}
→ create_ticket({"category":"defect","customer_id":"C-201","email_id":"E-1","priority":"normal","summary":"Router bought 3 November no longer switches on; customer asks for repair or replacement."})
← {"ticket_id": "T-901", "status": "created"}
→ create_ticket({"category":"billing","customer_id":"C-305","email_id":"E-3","priority":"normal","summary":"Invoice FV/11/2026 shows 4,200 PLN but the quote was 3,800 PLN; customer asks to correct it."})
← {"ticket_id": "T-902", "status": "created"}
→ create_ticket({"category":"delivery","customer_id":"C-118","email_id":"E-5","priority":"high","summary":"Order from 10 November was due yesterday but has not arrived and production line has stopped."})
← {"ticket_id": "T-903", "status": "created"}

Answer:

Summary of the triage:

**Tickets created**
- E-1 — Budmax (C-201, [email protected]) — defect, normal — T-901: router stopped working after 2 weeks (bought 3 Nov)
- E-3 — Ortopedix (C-305, [email protected]) — billing, normal — T-902: invoice FV/11/2026 amount mismatch (4,200 PLN vs 3,800 PLN quote)
- E-5 — Kartonpol (C-118, [email protected]) — delivery, **high** — T-903: order from 10 Nov missing, production line stopped

**Skipped (not complaints)**
- E-2 — newsletter (Tech Weekly #48) — skipped
- E-4 — spam/scam ("You have won an iPhone!") — skipped

**Not in customer system (no ticket created)**
- E-6 — [email protected] ("Printer complaint", prints stripes, wants return) — sender not found in the customer system; needs a person to handle/verify before a ticket can be raised.

No other senders were found besides these.

Qwen3.8-27B, 1 of 7 checks

Calls and returns:

→ list_inbox({"arguments": "[]"})
← {"error": "bad call: make.<locals>.list_inbox() got an unexpected keyword argument 'arguments'"}
→ list_inbox({"arguments": "{}"})
← {"error": "bad call: make.<locals>.list_inbox() got an unexpected keyword argument 'arguments'"}
→ list_inbox({"arguments": "null"})
← {"error": "bad call: make.<locals>.list_inbox() got an unexpected keyword argument 'arguments'"}
→ list_inbox({"arguments": "{\"x\":null}"})
← {"error": "bad call: make.<locals>.list_inbox() got an unexpected keyword argument 'arguments'"}
→ list_inbox({"arguments": "[]"})
← {"error": "bad call: make.<locals>.list_inbox() got an unexpected keyword argument 'arguments'"}
→ list_inbox({"arguments": "{\n  \"dummy\": \"\"\n}"})
← {"error": "bad call: make.<locals>.list_inbox() got an unexpected keyword argument 'arguments'"}
→ list_inbox({"arguments": "{}"})
← {"error": "bad call: make.<locals>.list_inbox() got an unexpected keyword argument 'arguments'"}
→ list_inbox({"arguments": "{\"id\": \"x\"}"})
← {"error": "bad call: make.<locals>.list_inbox() got an unexpected keyword argument 'arguments'"}
→ list_inbox({"arguments": "[]"})
← {"error": "bad call: make.<locals>.list_inbox() got an unexpected keyword argument 'arguments'"}
→ list_inbox({"arguments": "null"})
← {"error": "bad call: make.<locals>.list_inbox() got an unexpected keyword argument 'arguments'"}
→ list_inbox({"arguments": "null"})
← {"error": "bad call: make.<locals>.list_inbox() got an unexpected keyword argument 'arguments'"}
→ list_inbox({"arguments": "{\"id\":\"1\"}"})
← {"error": "bad call: make.<locals>.list_inbox() got an unexpected keyword argument 'arguments'"}
→ list_inbox({"arguments": "{\"a\":\"b\"}"})
← {"error": "bad call: make.<locals>.list_inbox() got an unexpected keyword argument 'arguments'"}
→ list_inbox({"arguments": "[]"})
← {"error": "bad call: make.<locals>.list_inbox() got an unexpected keyword argument 'arguments'"}
→ list_inbox({"arguments": "null"})
← {"error": "bad call: make.<locals>.list_inbox() got an unexpected keyword argument 'arguments'"}
→ list_inbox({"arguments": "{}"})
← {"error": "bad call: make.<locals>.list_inbox() got an unexpected keyword argument 'arguments'"}

No answer: the run ended without one.

Arete-27B, 1 of 7 checks

Calls and returns:

→ list_inbox({"emily": ")
← {"error": "bad call: Unterminated string starting at: line 1 column 11 (char 10)"}
→ list_inbox({"void": "true"})
← {"error": "bad call: make.<locals>.list_inbox() got an unexpected keyword argument 'void'"}
→ list_inbox({"arguments": "{}"})
← {"error": "bad call: make.<locals>.list_inbox() got an unexpected keyword argument 'arguments'"}
→ list_inbox({"arguments": "[]"})
← {"error": "bad call: make.<locals>.list_inbox() got an unexpected keyword argument 'arguments'"}
→ list_inbox({"arguments": ""})
← {"error": "bad call: make.<locals>.list_inbox() got an unexpected keyword argument 'arguments'"}
→ list_inbox({"arguments": "true"})
← {"error": "bad call: make.<locals>.list_inbox() got an unexpected keyword argument 'arguments'"}
→ list_inbox({"arguments": "{}"})
← {"error": "bad call: make.<locals>.list_inbox() got an unexpected keyword argument 'arguments'"}
→ list_inbox({"arguments": "{}"})
← {"error": "bad call: make.<locals>.list_inbox() got an unexpected keyword argument 'arguments'"}
→ list_inbox({"arguments": "{\"\": \"\"}"})
← {"error": "bad call: make.<locals>.list_inbox() got an unexpected keyword argument 'arguments'"}
→ list_inbox({"arguments": "null"})
← {"error": "bad call: make.<locals>.list_inbox() got an unexpected keyword argument 'arguments'"}
→ list_inbox({})
← {"emails": [{"id": "E-1", "from": "[email protected]", "subject": "Router stopped working after 2 weeks"}, {"id": "E-2", "from": "[email protected]", "subject": "Tech Weekly #48: this week's news"}, {"id": "E-3", "from": "[email protected]", "subject": "Wrong amount on invoice FV/11/2026"}, {"id": "E-4", "from": "[email protected]", "subject": "You have won an iPhone!"}, {"id": "E-5", "from": "[email protected]", "subject": "Delivery missing - our production has stopped"}, {"id": "E-6", "from": "[email protected]", "subject": "Printer complaint"}]}
→ read_email({"id": "E-1"})
← {"id": "E-1", "from": "[email protected]", "subject": "Router stopped working after 2 weeks", "body": "Hello, the router we bought on 3 November no longer switches on. Please repair or replace it. Jan Nowak, Budmax"}
→ read_email({"id": "E-2"})
← {"id": "E-2", "from": "[email protected]", "subject": "Tech Weekly #48: this week's news", "body": "This week: new processors, a laptop review and 10 security tips. Unsubscribe here."}
→ read_email({"id": "E-3"})
← {"id": "E-3", "from": "[email protected]", "subject": "Wrong amount on invoice FV/11/2026", "body": "Invoice FV/11/2026 says 4,200 PLN but the quote was 3,800 PLN. Please correct it. Ortopedix accounts"}
→ read_email({"id": "E-4"})
← {"id": "E-4", "from": "[email protected]", "subject": "You have won an iPhone!", "body": "Congratulations! Click the link and enter your card details to claim your prize."}
→ read_email({"id": "E-5"})
← {"id": "E-5", "from": "[email protected]", "subject": "Delivery missing - our production has stopped", "body": "The order from 10 November was due yesterday. It hasn't arrived and our production line has stopped. Urgent! Kartonpol"}

No answer: the run ended without one.

Meeting and room, guided prompt, English

The request:

Book a one-hour project meeting with Ola, Piotr and Marek on 24 November 2026, at the earliest time all three are free, between 09:00 and 17:00. Book a room for the three of them, then tell me the time and the room.

Inkling 2 Mini, 6 of 6 checks

Calls and returns:

→ get_busy({"date":"2026-11-24","person":"Ola"})
← {"person": "Ola", "date": "2026-11-24", "busy": ["09:00-10:30", "12:00-13:00"]}
→ get_busy({"date":"2026-11-24","person":"Piotr"})
← {"person": "Piotr", "date": "2026-11-24", "busy": ["10:00-11:00", "13:30-15:00"]}
→ get_busy({"date":"2026-11-24","person":"Marek"})
← {"person": "Marek", "date": "2026-11-24", "busy": ["09:00-09:30", "11:00-12:00", "15:00-16:00"]}
→ find_room({"date":"2026-11-24","end":"17:00","people":3,"start":"16:00"})
← {"date": "2026-11-24", "start": "16:00", "end": "17:00", "rooms": [{"room": "Sala B", "seats": 4}]}
→ create_event({"attendees":["Ola","Piotr","Marek"],"date":"2026-11-24","end":"17:00","room":"Sala B","start":"16:00","title":"Project meeting"})
← {"event_id": "EV-401", "status": "created"}

Answer:

Done. The earliest free hour for all three on 24 November 2026 is 16:00–17:00, booked in Sala B (4 seats). Attendees: Ola, Piotr, Marek. Event ID: EV-401.

Qwen3.8-27B, 6 of 6 checks

Calls and returns:

→ get_busy({"person": "Ola", "date": "2026-11-24"})
← {"person": "Ola", "date": "2026-11-24", "busy": ["09:00-10:30", "12:00-13:00"]}
→ get_busy({"person": "Piotr", "date": "2026-11-24"})
← {"person": "Piotr", "date": "2026-11-24", "busy": ["10:00-11:00", "13:30-15:00"]}
→ get_busy({"person": "Marek", "date": "2026-11-24"})
← {"person": "Marek", "date": "2026-11-24", "busy": ["09:00-09:30", "11:00-12:00", "15:00-16:00"]}
→ find_room({"date": "2026-11-24", "start": "16:00", "end": "17:00", "people": 3})
← {"date": "2026-11-24", "start": "16:00", "end": "17:00", "rooms": [{"room": "Sala B", "seats": 4}]}
→ create_event({"title": "Project meeting", "date": "2026-11-24", "start": "16:00", "end": "17:00", "attendees": ["Ola", "Piotr", "Marek"], "room": "Sala B"})
← {"event_id": "EV-401", "status": "created"}

Answer:

Meeting booked:

- **Date:** 24 November 2026
- **Time:** 16:00–17:00
- **Room:** Sala B
- **Attendees:** Ola, Piotr, Marek

This was the earliest one-hour slot on that day when all three were free.

Arete-27B, 6 of 6 checks

Calls and returns:

→ get_busy({"person": "Ola", "date": "2026-11-24"})
← {"person": "Ola", "date": "2026-11-24", "busy": ["09:00-10:30", "12:00-13:00"]}
→ get_busy({"person": "Piotr", "date": "2026-11-24"})
← {"person": "Piotr", "date": "2026-11-24", "busy": ["10:00-11:00", "13:30-15:00"]}
→ get_busy({"person": "Marek", "date": "2026-11-24"})
← {"person": "Marek", "date": "2026-11-24", "busy": ["09:00-09:30", "11:00-12:00", "15:00-16:00"]}
→ find_room({"date": "2026-11-24", "start": "16:00", "end": "17:00", "people": 3})
← {"date": "2026-11-24", "start": "16:00", "end": "17:00", "rooms": [{"room": "Sala B", "seats": 4}]}
→ create_event({"title": "Project Meeting", "date": "2026-11-24", "start": "16:00", "end": "17:00", "attendees": ["Ola", "Piotr", "Marek"], "room": "Sala B"})
← {"event_id": "EV-401", "status": "created"}

Answer:

The one-hour project meeting is booked for **24 November 2026, 16:00–17:00** (the earliest time all three are free) in **Sala B**, with **Ola, Piotr and Marek** attending.

Order with a short item, short prompt, English

The request:

Place an order for client K-7: 10 × A-100 and 5 × B-200. If an item isn't available in the full quantity, order its recommended substitute instead (the whole quantity). Then tell me the order number, what was ordered and the total.

Inkling 2 Mini, 2 of 7 checks

Calls and returns:

→ check_stock({"sku":"A-100"})
← {"sku": "A-100", "name": "HDMI cable 2 m", "available": 50, "price": 24.9}
→ check_stock({"sku":"B-200"})
← {"sku": "B-200", "name": "65 W laptop charger", "available": 2, "price": 119.0}
→ find_substitute({"sku":"B-200"})
← {"sku": "B-200", "substitute": "B-210", "name": "65 W USB-C charger (substitute)"}
→ check_stock({"sku":"B-210"})
← {"sku": "B-210", "name": "65 W USB-C charger (substitute)", "available": 20, "price": 125.0}
→ place_order({"client_id":"K-7","items":[{"qty":10,"sku":"A-100"},{"qty":5,"sku":"B-210"}]})
← {"error": "price_changed", "sku": "B-210", "old_price": 125.0, "new_price": 129.0, "message": "The price of B-210 has changed to 129.00 since the last quote. Repeat the order with confirm_prices=true to accept it."}

Answer:

I ran into a price change before placing the order, so I haven't confirmed yet.

Here's where things stand:

- **A-100** (HDMI cable 2 m) — 10 × 24.90 available (50 in stock): OK to order as requested.
- **B-200** (65 W laptop charger) — only 2 in stock, need 5. Recommended substitute is **B-210** (65 W USB-C charger), which is in stock (20 available) — however, its price has changed from 125.00 to **129.00** since the last quote.

If I place the order as-is with the new price, the total would be 10 × 24.90 + 5 × 129.00 = **874.00 → wait, recalculating: 249.00 + 645.00 = 894.00**.

Do you want me to confirm and place the order with the updated B-210 price (129.00), or do you want something else?

Qwen3.8-27B, 7 of 7 checks

Calls and returns:

→ check_stock({"sku": "A-100"})
← {"sku": "A-100", "name": "HDMI cable 2 m", "available": 50, "price": 24.9}
→ check_stock({"sku": "B-200"})
← {"sku": "B-200", "name": "65 W laptop charger", "available": 2, "price": 119.0}
→ find_substitute({"sku": "B-200"})
← {"sku": "B-200", "substitute": "B-210", "name": "65 W USB-C charger (substitute)"}
→ check_stock({"sku": "B-210"})
← {"sku": "B-210", "name": "65 W USB-C charger (substitute)", "available": 20, "price": 125.0}
→ place_order({"client_id": "K-7", "items": [{"sku": "A-100", "qty": 10}, {"sku": "B-210", "qty": 5}]})
← {"error": "price_changed", "sku": "B-210", "old_price": 125.0, "new_price": 129.0, "message": "The price of B-210 has changed to 129.00 since the last quote. Repeat the order with confirm_prices=true to accept it."}
→ place_order({"client_id": "K-7", "items": [{"sku": "A-100", "qty": 10}, {"sku": "B-210", "qty": 5}], "confirm_prices": true})
← {"order_id": "ZAM-5531", "status": "placed", "items": [{"sku": "A-100", "qty": 10, "price": 24.9}, {"sku": "B-210", "qty": 5, "price": 129.0}], "total": 894.0}

Answer:

Order placed successfully. ✅

- **Order number:** ZAM-5531
- **Client:** K-7
- **Items ordered:**
  - 10 × A-100 (HDMI cable 2 m) @ 24.90 = 249.00
  - 5 × B-210 (65 W USB-C charger – substitute for B-200) @ 129.00 = 645.00

**Total: 894.00**

Notes:
- B-200 (65 W laptop charger) only had 2 units available, so I substituted the full quantity of 5 with the recommended substitute B-210 (20 available).
- The price of B-210 changed from 125.00 to 129.00 at order time; I accepted the new price, which is why the total is 894.00.

Arete-27B, 7 of 7 checks

Calls and returns:

→ check_stock({"sku": "A-100"})
← {"sku": "A-100", "name": "HDMI cable 2 m", "available": 50, "price": 24.9}
→ check_stock({"sku": "B-200"})
← {"sku": "B-200", "name": "65 W laptop charger", "available": 2, "price": 119.0}
→ find_substitute({"sku": "B-200"})
← {"sku": "B-200", "substitute": "B-210", "name": "65 W USB-C charger (substitute)"}
→ check_stock({"sku": "B-210"})
← {"sku": "B-210", "name": "65 W USB-C charger (substitute)", "available": 20, "price": 125.0}
→ place_order({"client_id": "K-7", "items": [{"sku": "A-100", "qty": 10}, {"sku": "B-210", "qty": 5}]})
← {"error": "price_changed", "sku": "B-210", "old_price": 125.0, "new_price": 129.0, "message": "The price of B-210 has changed to 129.00 since the last quote. Repeat the order with confirm_prices=true to accept it."}
→ place_order({"client_id": "K-7", "items": [{"qty": 10, "sku": "A-100"}, {"qty": 5, "sku": "B-210"}], "confirm_prices": true})
← {"order_id": "ZAM-5531", "status": "placed", "items": [{"sku": "A-100", "qty": 10, "price": 24.9}, {"sku": "B-210", "qty": 5, "price": 129.0}], "total": 894.0}

Answer:

Order placed successfully. Here's the summary for client **K-7**:

**Order number:** ZAM-5531

**Items ordered:**
- **A-100** — HDMI cable 2 m: **10 pcs** @ 24.90 = 249.00 (fully available, 50 in stock)
- **B-210** — 65 W USB-C charger (substitute): **5 pcs** @ 129.00 = 645.00
  - *Note:* B-200 (65 W laptop charger) only had 2 in stock, so the full quantity of 5 was ordered as its recommended substitute B-210. Its price was updated from 125.00 to 129.00, which was accepted on the order.

**Total:** **894.00**

Overdue invoices (Polish), guided prompt

The request:

Sprawdź, które faktury klienta Hydro-Tech Sp. z o.o. są na dziś po terminie płatności i ile z każdej pozostało do zapłaty. Policz łączną zaległą kwotę i liczbę dni opóźnienia każdej faktury. Potem przygotuj uprzejmy e-mail z przypomnieniem o płatności do osoby kontaktowej klienta (numery faktur, kwoty do zapłaty, dni opóźnienia, suma) i zapisz go jako szkic. Na koniec napisz mi krótko, co zrobiłeś.

Inkling 2 Mini, 13 of 13 checks

Calls and returns:

→ szukaj_klienta({"nazwa":"Hydro-Tech Sp. z o.o."})
← {"wyniki": [{"id": "K-114", "nazwa": "Hydro-Tech Sp. z o.o.", "nip": "6762498310", "email": "[email protected]", "osoba_kontaktowa": "Anna Wiśniewska (główna księgowa)"}, {"id": "K-201", "nazwa": "Hydrotech Serwis s.c.", "nip": "9452173006", "email": "[email protected]", "osoba_kontaktowa": "Marek Zając"}]}
→ lista_faktur({"klient_id":"K-114"})
← {"klient_id": "K-114", "faktury": [{"numer": "FV/2026/08/077", "data_wystawienia": "2026-08-14", "termin_platnosci": "2026-09-13", "kwota_brutto": 1845.0, "zaplacono": 1845.0}, {"numer": "FV/2026/09/031", "data_wystawienia": "2026-09-10", "termin_platnosci": "2026-10-10", "kwota_brutto": 12300.0, "zaplacono": 12300.0}, {"numer": "FV/2026/09/044", "data_wystawienia": "2026-09-22", "termin_platnosci": "2026-10-22", "kwota_brutto": 4920.0, "zaplacono": 0.0}, {"numer": "FV/2026/10/012", "data_wystawienia": "2026-10-05", "termin_platnosci": "2026-11-04", "kwota_brutto": 8610.0, "zaplacono": 3000.0}, {"numer": "FV/2026/10/058", "data_wystawienia": "2026-10-28", "termin_platnosci": "2026-11-27", "kwota_brutto": 2460.0, "zaplacono": 0.0}]}
→ oblicz({"wyrazenie":"4920.00 - 0.00"})
← {"wynik": 4920.0}
→ oblicz({"wyrazenie":"8610.00 - 3000.00"})
← {"wynik": 5610.0}
→ oblicz({"wyrazenie":"2460.00 - 0.00"})
← {"wynik": 2460.0}
→ oblicz({"wyrazenie":"1845.00 - 1845.00"})
← {"wynik": 0.0}
→ oblicz({"wyrazenie":"12300.00 - 12300.00"})
← {"wynik": 0.0}
→ dni_miedzy({"data_do":"2026-11-19","data_od":"2026-10-22"})
← {"dni": 28}
→ dni_miedzy({"data_do":"2026-11-19","data_od":"2026-11-04"})
← {"dni": 15}
→ oblicz({"wyrazenie":"4920.00 + 5610.00"})
← {"wynik": 10530.0}
→ zapisz_szkic({"do":"[email protected]","temat":"Przypomnienie o płatności – zaległe faktury","tresc":"Szanowna Pani Anno,\n\nprzypominamy o poniższych faktuarach, które są na dziś po terminie płatności:\n\n- FV/2026/09/044 – kwota do zapłaty: 4 920,00 zł – 28 dni opóźnienia\n- FV/2026/10/012 – kwota do zapłaty: 5 610,00 zł – 15 dni opóźnienia\n\nŁączna zaległa kwota: 10 530,00 zł\n\nProsimy o uregulowanie należności w możliwie najkrótszym terminie. W razie pytań pozostajemy do dyspozycji.\n\nZ poważaniem,\nKowalski Serwis Sp. z o.o."})
← {"status": "zapisano", "id_szkicu": "SZ-1"}

The draft it saved:

Do: [email protected]
Temat: Przypomnienie o płatności – zaległe faktury

Szanowna Pani Anno,

przypominamy o poniższych faktuarach, które są na dziś po terminie płatności:

- FV/2026/09/044 – kwota do zapłaty: 4 920,00 zł – 28 dni opóźnienia
- FV/2026/10/012 – kwota do zapłaty: 5 610,00 zł – 15 dni opóźnienia

Łączna zaległa kwota: 10 530,00 zł

Prosimy o uregulowanie należności w możliwie najkrótszym terminie. W razie pytań pozostajemy do dyspozycji.

Z poważaniem,
Kowalski Serwis Sp. z o.o.

Answer:

Gotowe. Podsumowanie:

- Klient: Hydro-Tech Sp. z o.o. (K-114), kontakt: Anna Wiśniewska ([email protected])
- Po terminie płatności (dziś: 2026-11-19):
  - FV/2026/09/044 – do zapłaty 4 920,00 zł – 28 dni opóźnienia (termin 2026-10-22)
  - FV/2026/10/012 – do zapłaty 5 610,00 zł – 15 dni opóźnienia (termin 2026-11-04)
- Łączna zaległa kwota: 10 530,00 zł
- Pozostałe faktury: FV/2026/08/077 i FV/2026/09/031 opłacone w całości (0,00 zł do zapłaty); FV/2026/10/058 (2 460,00 zł) nie jest jeszcze po terminie – termin 2026-11-27, więc nie ujęta w przypomnieniu.
- E-mail z przypomnieniem zapisany jako szkic (id_szkicu: SZ-1) – nie został wysłany.

Qwen3.8-27B, 13 of 13 checks

Calls and returns:

→ szukaj_klienta({"nazwa": "Hydro-Tech"})
← {"wyniki": [{"id": "K-114", "nazwa": "Hydro-Tech Sp. z o.o.", "nip": "6762498310", "email": "[email protected]", "osoba_kontaktowa": "Anna Wiśniewska (główna księgowa)"}, {"id": "K-201", "nazwa": "Hydrotech Serwis s.c.", "nip": "9452173006", "email": "[email protected]", "osoba_kontaktowa": "Marek Zając"}]}
→ lista_faktur({"klient_id": "K-114"})
← {"klient_id": "K-114", "faktury": [{"numer": "FV/2026/08/077", "data_wystawienia": "2026-08-14", "termin_platnosci": "2026-09-13", "kwota_brutto": 1845.0, "zaplacono": 1845.0}, {"numer": "FV/2026/09/031", "data_wystawienia": "2026-09-10", "termin_platnosci": "2026-10-10", "kwota_brutto": 12300.0, "zaplacono": 12300.0}, {"numer": "FV/2026/09/044", "data_wystawienia": "2026-09-22", "termin_platnosci": "2026-10-22", "kwota_brutto": 4920.0, "zaplacono": 0.0}, {"numer": "FV/2026/10/012", "data_wystawienia": "2026-10-05", "termin_platnosci": "2026-11-04", "kwota_brutto": 8610.0, "zaplacono": 3000.0}, {"numer": "FV/2026/10/058", "data_wystawienia": "2026-10-28", "termin_platnosci": "2026-11-27", "kwota_brutto": 2460.0, "zaplacono": 0.0}]}
→ oblicz({"wyrazenie": "1845.0 - 1845.0"})
← {"wynik": 0.0}
→ oblicz({"wyrazenie": "12300.0 - 12300.0"})
← {"wynik": 0.0}
→ oblicz({"wyrazenie": "4920.0 - 0.0"})
← {"wynik": 4920.0}
→ oblicz({"wyrazenie": "8610.0 - 3000.0"})
← {"wynik": 5610.0}
→ oblicz({"wyrazenie": "2460.0 - 0.0"})
← {"wynik": 2460.0}
→ dni_miedzy({"data_od": "2026-10-22", "data_do": "2026-11-19"})
← {"dni": 28}
→ dni_miedzy({"data_do": "2026-11-19", "data_od": "2026-11-04"})
← {"dni": 15}
→ oblicz({"wyrazenie": "4920.0 + 5610.0"})
← {"wynik": 10530.0}
→ zapisz_szkic({"do": "[email protected]", "temat": "Przypomnienie o płatności – faktury FV/2026/09/044 i FV/2026/10/012 (Hydro-Tech Sp. z o.o.)", "tresc": "Szanowna Pani Anno,\n\nuprzejmie przypominamy o rozliczeniu poniższych faktur, których termin płatności minął na dzień 19 listopada 2026 r.:\n\n1. Faktura FV/2026/09/044 z dnia 22.09.2026 – termin płatności: 22.10.2026, kwota do zapłaty: 4920,00 zł, opóźnienie: 28 dni.\n2. Faktura FV/2026/10/012 z dnia 05.10.2026 – termin płatności: 04.11.2026, kwota do zapłaty: 5610,00 zł, opóźnienie: 15 dni.\n\nŁączna kwota zaległości: 10530,00 zł.\n\nProsimy o uregulowanie powyższych należności najszybciej jak to możliwe. W razie jakichkolwiek pytań dotyczących faktur lub formy rozliczenia pozostajemy do dyspozycji.\n\nZ poważaniem,\nKowalski Serwis Sp. z o.o."})
← {"status": "zapisano", "id_szkicu": "SZ-1"}

The draft it saved:

Do: [email protected]
Temat: Przypomnienie o płatności – faktury FV/2026/09/044 i FV/2026/10/012 (Hydro-Tech Sp. z o.o.)

Szanowna Pani Anno,

uprzejmie przypominamy o rozliczeniu poniższych faktur, których termin płatności minął na dzień 19 listopada 2026 r.:

1. Faktura FV/2026/09/044 z dnia 22.09.2026 – termin płatności: 22.10.2026, kwota do zapłaty: 4920,00 zł, opóźnienie: 28 dni.
2. Faktura FV/2026/10/012 z dnia 05.10.2026 – termin płatności: 04.11.2026, kwota do zapłaty: 5610,00 zł, opóźnienie: 15 dni.

Łączna kwota zaległości: 10530,00 zł.

Prosimy o uregulowanie powyższych należności najszybciej jak to możliwe. W razie jakichkolwiek pytań dotyczących faktur lub formy rozliczenia pozostajemy do dyspozycji.

Z poważaniem,
Kowalski Serwis Sp. z o.o.

Answer:

Gotowe. Krótkie podsumowanie:

- **Klient:** Hydro-Tech Sp. z o.o. (K-114), kontakt: Anna Wiśniewska, e-mail: [email protected].
- **Faktury po terminie (stan na 19.11.2026):**
  - **FV/2026/09/044** – termin 22.10.2026, do zapłaty **4 920,00 zł**, opóźnienie **28 dni**.
  - **FV/2026/10/012** – termin 04.11.2026, do zapłaty **5 610,00 zł** (zapłacono 3 000,00 zł z 8 610,00 zł), opóźnienie **15 dni**.
- **Łączna zaległość: 10 530,00 zł.**
- Pozostałe faktury: FV/2026/08/077 i FV/2026/09/031 są w pełni opłacone, a FV/2026/10/058 (2 460,00 zł) ma termin 27.11.2026 i jeszcze nie jest po terminie.
- Uprzejmий e-mail z przypomnieniem (numery faktur, kwoty, dni opóźnienia, suma) został zapisany jako **szkic SZ-1** do adresata [email protected] — nie został wysłany, czeka na zatwierdzenie.

Arete-27B, 13 of 13 checks

Calls and returns:

→ szukaj_klienta({"nazwa": "Hydro-Tech"})
← {"wyniki": [{"id": "K-114", "nazwa": "Hydro-Tech Sp. z o.o.", "nip": "6762498310", "email": "[email protected]", "osoba_kontaktowa": "Anna Wiśniewska (główna księgowa)"}, {"id": "K-201", "nazwa": "Hydrotech Serwis s.c.", "nip": "9452173006", "email": "[email protected]", "osoba_kontaktowa": "Marek Zając"}]}
→ lista_faktur({"klient_id": "K-114"})
← {"klient_id": "K-114", "faktury": [{"numer": "FV/2026/08/077", "data_wystawienia": "2026-08-14", "termin_platnosci": "2026-09-13", "kwota_brutto": 1845.0, "zaplacono": 1845.0}, {"numer": "FV/2026/09/031", "data_wystawienia": "2026-09-10", "termin_platnosci": "2026-10-10", "kwota_brutto": 12300.0, "zaplacono": 12300.0}, {"numer": "FV/2026/09/044", "data_wystawienia": "2026-09-22", "termin_platnosci": "2026-10-22", "kwota_brutto": 4920.0, "zaplacono": 0.0}, {"numer": "FV/2026/10/012", "data_wystawienia": "2026-10-05", "termin_platnosci": "2026-11-04", "kwota_brutto": 8610.0, "zaplacono": 3000.0}, {"numer": "FV/2026/10/058", "data_wystawienia": "2026-10-28", "termin_platnosci": "2026-11-27", "kwota_brutto": 2460.0, "zaplacono": 0.0}]}
→ oblicz({"wyrazenie": "1845.0 - 1845.0"})
← {"wynik": 0.0}
→ oblicz({"wyrazenie": "12300.0 - 12300.0"})
← {"wynik": 0.0}
→ oblicz({"wyrazenie": "4920.0 - 0.0"})
← {"wynik": 4920.0}
→ oblicz({"wyrazenie": "8610.0 - 3000.0"})
← {"wynik": 5610.0}
→ oblicz({"wyrazenie": "2460.0 - 0.0"})
← {"wynik": 2460.0}
→ dni_miedzy({"data_od": "2026-10-22", "data_do": "2026-11-19"})
← {"dni": 28}
→ dni_miedzy({"data_od": "2026-11-04", "data_do": "2026-11-19"})
← {"dni": 15}
→ oblicz({"wyrazenie": "4920.0 + 5610.0"})
← {"wynik": 10530.0}
→ zapisz_szkic({"do": "[email protected]", "temat": "Przypomnienie o płatności – faktury po terminie (FV/2026/09/044, FV/2026/10/012)", "tresc": "Szanowna Pani Anno,\n\nzwracam się z uprzejmą prośbą o sprawdzenie i uregulowanie faktur, których terminy płatności już upłynęły. Na dzień 19.11.2026 r. następujące dokumenty są po terminie:\n\n- Faktura FV/2026/09/044 (termin płatności: 22.10.2026) – kwota do zapłaty: 4 920,00 zł brutto, opóźnienie: 28 dni.\n- Faktura FV/2026/10/012 (termin płatności: 04.11.2026) – kwota do zapłaty: 5 610,00 zł brutto, opóźnienie: 15 dni.\n\nŁączna zaległość wynosi: 10 530,00 zł brutto.\n\nW przypadku dokonanych już wpłat lub planowanych płatności proszę o informację zwrotną. W razie jakichkolwiek pytań pozostaję do dyspozycji.\n\nZ poważaniem,\nKowalski Serwis Sp. z o.o."})
← {"status": "zapisano", "id_szkicu": "SZ-1"}

The draft it saved:

Do: [email protected]
Temat: Przypomnienie o płatności – faktury po terminie (FV/2026/09/044, FV/2026/10/012)

Szanowna Pani Anno,

zwracam się z uprzejmą prośbą o sprawdzenie i uregulowanie faktur, których terminy płatności już upłynęły. Na dzień 19.11.2026 r. następujące dokumenty są po terminie:

- Faktura FV/2026/09/044 (termin płatności: 22.10.2026) – kwota do zapłaty: 4 920,00 zł brutto, opóźnienie: 28 dni.
- Faktura FV/2026/10/012 (termin płatności: 04.11.2026) – kwota do zapłaty: 5 610,00 zł brutto, opóźnienie: 15 dni.

Łączna zaległość wynosi: 10 530,00 zł brutto.

W przypadku dokonanych już wpłat lub planowanych płatności proszę o informację zwrotną. W razie jakichkolwiek pytań pozostaję do dyspozycji.

Z poważaniem,
Kowalski Serwis Sp. z o.o.

Answer:

Gotowe. Oto co zrobiłem:

1. **Znalazłem klienta**: Hydro-Tech Sp. z o.o. (K-114, NIP 6762498310), kontakt: Anna Wiśniewska, e-mail: [email protected].
2. **Sprawdziłem 5 faktur** i policzyłem kwoty do zapłaty:
   - FV/2026/08/077 i FV/2026/09/031 — opłacone w całości (0,00 zł).
   - **FV/2026/09/044** — termin 22.10.2026, **4 920,00 zł** do zapłaty, **28 dni** opóźnienia.
   - **FV/2026/10/012** — termin 04.11.2026, **5 610,00 zł** do zapłaty (zapłacono 3 000,00 zł), **15 dni** opóźnienia.
   - FV/2026/10/058 — 2 460,00 zł, ale termin 27.11.2026 jeszcze nie minął.
3. **Łączna zaległość: 10 530,00 zł brutto**.
4. **Zapiszłem szkic e-maila** (id SZ-1) do Pani Anny Wiśniewskiej z uprzejmym przypomnieniem, zawierającym numery faktur, kwoty, dni opóźnienia i łączną kwotę. Szkic czeka na wysłanie przez człowieka.

Sources

  1. Thinking Machines Lab (2026). Inkling 2 Mini, model card, revision b85c3cf. https://huggingface.co/thinkingmachines/Inkling-2-Mini
  2. Qwen Team (2026). Qwen3.8-27B-FP8, model card, revision c3a462c. https://huggingface.co/Qwen/Qwen3.8-27B-FP8
  3. supz (2026). Arete-27B-W4A16-AutoRound, model card, revision 67b0699. https://huggingface.co/supz/Arete-27B-W4A16-AutoRound
  4. vLLM project (2026). [Bug]: Inkling 2 Mini Docker image missing H100 CUDA kernels, issue 69046 (read 22 Nov 2026). https://github.com/vllm-project/vllm/issues/69046
  5. Grace, M., Hadfield, J., Olivares, R. and De Jonghe, J. (2026). Demystifying evals for AI agents. Anthropic, 9 January 2026. https://www.anthropic.com/engineering/demystifying-evals-for-ai-agents
  6. NIST (2023). Artificial Intelligence Risk Management Framework (AI RMF 1.0), NIST AI 100-1, section 5, the core. https://airc.nist.gov/airmf-resources/airmf/5-sec-core/
  7. Artificial Analysis (2026). Intelligence Index v5.0.1, models leaderboard, and the pages of Inkling 2 Mini and Qwen3.8 27B (read 22 Nov 2026). https://artificialanalysis.ai/leaderboards/models
  8. Hugging Face (2026). Download counts of thinkingmachines/Inkling-2-Mini, Qwen/Qwen3.8-27B and NousResearch/Arete-27B, and the weight precision of the last two, from its model API (read 22 Nov 2026). https://huggingface.co/thinkingmachines/Inkling-2-Mini
  9. OpenRouter (2026). Models list and the endpoints of Inkling 2 Mini and Qwen3.8-27B (read 22 Nov 2026). https://openrouter.ai/models
  10. Thinking Machines Lab (2026). Introducing Inkling 2 Mini, 20 November 2026, section "NVFP4 and low-bit quantization". https://thinkingmachines.ai/news/inkling-2-mini
  11. NVIDIA (2026). Qwen3.8-27B-NVFP4, model card, revision 482ca0f. https://huggingface.co/nvidia/Qwen3.8-27B-NVFP4
📰 Read the original article on Dev.to AI

Originally published by Dev.to AI. Aggregated on AIWithGhost for educational purposes — full credit and traffic to the original publisher.