I ran a 177B open model on Colab G4 - here is the exact workflow, and the prompt I run before downloading anything

Last week I ran a 177B-parameter open model (Qwen3.8-Flash-Next, abliterated IQ3_S build) on a Colab G4 runtime. No local GPU, no API bill. This post is the exact workflow, with the numbers I actually measured - and the one prompt I now run before I download 80+ GB of model weights.

Measured on 2026-11-09, not promised for your run:

• Runtime: Colab G4, RTX PRO 6000, 96 GB VRAM
• Model file: 83.86 GB (IQ3_S quantization)
• Download time: 4 min 13 s
• Context: 32768 (-c 32768 on llama-server)
• First chat reply: 0.9 s
• Cost at test time: about 10.05 compute units/hour on Google AI Pro (200 CU/month)

I did not measure steady-state tokens/second, so I won't quote one. A server that starts is not the same as a server that answers well.

Step 0: Ask before you download

The expensive mistake is downloading a model that cannot run on your hardware. I use this prompt first (it is Prompt A1 from my pack, shown in full):

I want to run [MODEL NAME AND SIZE] on [PLATFORM AND RUNTIME]. My goal is [GOAL]. My constraints are: VRAM [X GB], RAM [X GB], disk [X GB], budget [X per hour], time limit [X hours]. Tell me the exact runtime, context size, quantization, and disk space I need. Then give me a 2-minute go / no-go check I can run before I download anything. Flag anything that would make this fail on my hardware.

Filled in for this run, it returned: G4 runtime, IQ3_S, 32k context, ~84 GB disk plus headroom, and a go/no-go check (nvidia-smi shows 96 GB, df shows enough free disk). It said go. It was right.

When something does fail, I paste the full error log into a second prompt that asks for the most likely cause in one sentence, three fixes ranked most-to-least likely with exact commands, how to verify, and what to check next. The "paste the full log, not a summary" part matters - the key line is usually near the end.

The workflow

1. Start a G4 runtime. The free T4 (15 GB) cannot run this model. Do not try to be clever here.
2. Download the GGUF (IQ3_S, 83.86 GB). About 4 minutes on the test run.
3. Start llama-server with -c 32768. Wait for the health check, not just the process.
4. Smoke test: one short prompt, confirm a sane reply (0.9 s for the first reply in my run), then start real work.
5. Save your chat history to Drive as you go - see the pitfall below. This is the step everyone skips and regrets.

The pitfall that cost me a session

Colab runtimes disconnect. When they do, /content is wiped: the model is gone (fine, re-download is 4 minutes) and your chat history is gone (not fine).

Saving the notebook file is not saving the runtime state. What works:

1. Mount Google Drive in the first cell, before anything else.
2. Add a save cell that writes chat_history.json to Drive (e.g. /content/drive/MyDrive/colab-chat/chat_history.json) and run it after every important exchange, not at the end. You will not get a warning before a disconnect.
3. Add a load cell that restores the history on a fresh runtime, and test it once on purpose: restart, load, confirm your messages are back.
4. Verify the file exists in the Drive web UI. In my first test run it did not - the save step had never run before the session ended, and the test chats were lost.

Do not try to save the 81 GB model itself to Drive to skip the re-download. It is too large and too slow. Accept the 4 minutes.

The full pack

This workflow is one of four in the pack I sell: AI Workflow Pack for Builders - 25 prompts + 4 tested workflows (running big open models, two free-tier VLESS node setups with real dial tests, and the Colab chat-saving workflow above). Every workflow in it was run by me in November 2026, with measured numbers labelled as test results, not promises.

• Free sample (3 prompts + 1 workflow): https://ko-fi.com/ytk821442
• Full pack, $9 (personal) / $19 (team of 5 + free v1.x updates): https://ko-fi.com/s/1946e2a8b2

If you try the go/no-go prompt on your own setup, I would like to hear what it flagged for you.