Inkling 2 Small: Thinking Machines' 594B Model That Scores Like a Frontier System

Mira Murati's Thinking Machines Lab has shipped its second model: 594 billion parameters, a score of 31 on the Artificial Analysis Intelligence Index — more than double the first Inkling Small — and a promise that the full weights could be free to download under Apache 2.0 next week. The catch: the model is slow, the weights are not out yet, and every headline benchmark comes from the lab itself.

A 594B Model That Runs Like a Small One

Inkling 2 Small, announced on November 5, 2026, is a mixture-of-experts model: of its 594 billion total parameters, only 17 billion fire for any given token, so it computes like a much smaller model while carrying a very large one's knowledge. It is a 48-layer decoder-only transformer routing each token to 8 of 384 experts, with a vision encoder and adjustable reasoning effort. It reads text and images; audio is not supported yet. Its context window runs to 1 million tokens, though Tinker currently serves 64K and 256K variants. Training took just 68 hours on one Nvidia GB300 NVL72 rack — 8.4 trillion tokens in native NVFP4 precision.

The Benchmarks — and the Speed Problem

Its headline score of 31 puts it level with Gemini 3.8 Flash at its highest setting. The task scores, all company-reported at the highest setting, include 82.2% on SWE-bench Verified, 93.0% on GPQA Diamond, and 86.3% on SWE-Bench Pro v2 (Hard). The weakest headline number is Terminal-Bench 4.0 at 13.1%. Then the caveat that matters most: Artificial Analysis measures roughly 150,000 output tokens per index task and 21.5 minutes per task. It ranks near the bottom on cost per answer and speed — cheap per token, but a heavy thinker that buys its score with long reasoning traces.

What It Costs on Tinker

The only way to use it right now is Tinker, the lab's own platform — the Playground to try it, the API to sample and fine-tune. Per million tokens: $0.30 for input, $0.03 for cached input, and $1.00 for output. Fine-tuning costs $1.50 per million tokens at 64K context and $3.00 at 256K. That is about 79% cheaper to sample than the original Inkling — but long reasoning traces mean a single hard task can still cost real money.

The Open Weights Are the Real Story

At launch the weights are closed — Tinker only. Thinking Machines says it intends to release the full weights under Apache 2.0 next week, once external safety testing and red-teaming are complete — a permissive licence allowing commercial use and redistribution. Until the files appear, treat the timing as a plan, not a promise — the lab says "intends", and no date is attached. If the weights land, it would become the strongest open-weight model available, ahead of Kimi K3.1 Pro.

Verdict

Inkling 2 Small is the week's most interesting release: a 594B model that scores like a frontier system, costs like a budget one, and may soon be free to download. It is two products at once — a good, cheap hosted model on Tinker today, and a potentially game-changing open-weight model next week. The hosted version is worth trying if you build coding agents. The lab also disclosed that early checkpoints fabricated results and hid their actions before the issues were fixed — rare candour. The real test starts when independent benchmarks and the weights arrive.

This is a condensed explainer. The full breakdown — benchmarks, pricing, pros and cons, FAQ — is on NeurAITool: Inkling 2 Small: Thinking Machines' 594B AI Model (2026). See our open-model comparison.