Thinking Machines Lab just shipped its second-generation AI model, and it is a big one in every sense. Inkling 2 Small packs 594 billion parameters, scores 31 on the Artificial Analysis Intelligence Index — more than double its predecessor — and, if the lab keeps its promise, its full weights will be free to download under Apache 2.0 next week. Here is what the model actually is, what it costs, and who should care.

What Is Inkling 2 Small?

Inkling 2 Small is the first model of Thinking Machines Lab's second generation, announced on November 5, 2026. The lab was founded by former OpenAI CTO Mira Murati, and its first models — Inkling and Inkling Small — arrived in July 2026. Inkling 2 Small is a mixture-of-experts (MoE) model: of its 594 billion total parameters, only 17 billion are active for each token. That means it computes roughly like a much smaller model while carrying the knowledge of a very large one.

The architecture is a 48-layer decoder-only transformer in which each token is routed to 8 of 384 experts, with a vision encoder for image input and support for adjustable reasoning effort. It reads text and images, but does not support audio yet. The model card lists a context window of up to 1 million tokens, although the hosted Tinker endpoint currently serves it in 64K and 256K variants.

The training run itself is part of the story. According to the lab, pretraining ran for just 68 hours on a single Nvidia GB300 NVL72 rack, processing 8.4 trillion tokens — trained natively in NVFP4 precision rather than converted afterwards. Post-training involved 17 million rollouts over 9 days. For a model of this size, that is an unusually fast training cycle, and it is one reason the release caught the AI world off guard.

How Good Is It? The Benchmarks

The headline number is 31 on the Artificial Analysis Intelligence Index, up from 13 for the original Inkling Small and level with Gemini 3.8 Flash at its highest setting. All benchmarks below are company-reported, run at the highest effort setting, so treat them as the lab's best-case numbers rather than independent results.

  • SWE-bench Verified: 82.2% — strong for real-world coding tasks.
  • GPQA Diamond: 93.0% — graduate-level science questions.
  • Humanity's Last Exam (text): 44.1%.
  • SWE-Bench Pro v2 (Hard): 86.3%.
  • DeepSWE v1.1: 72.1%.
  • Terminal-Bench 4.0: 13.1% — the weakest of the headline scores, and a reminder that terminal-agent work is still hard for every model.

There is one honest caveat worth knowing. Artificial Analysis measures roughly 150,000 output tokens per index task for this model, with about 21.5 minutes per task and 34.5 seconds to the first answer token. It ranks near the bottom of the index on cost per answer and on speed. Inkling 2 Small is cheap per token, but it is a heavy thinker — it buys its score with long, expensive reasoning traces. If you need fast answers, that matters more than the index score.

Pricing: What It Costs on Tinker

Right now, the only way to use Inkling 2 Small is through Tinker, the lab's own fine-tuning and inference platform. It is available in the Tinker Playground for trying it out, and for sampling and fine-tuning through the Tinker API. Pricing per million tokens:

  • Input: $0.30 per million tokens
  • Cached input: $0.03 per million tokens
  • Output: $1.00 per million tokens
  • Fine-tuning: $1.50 per million tokens at 64K context, $3.00 at 256K

That is about 79% cheaper to sample than the original Inkling, and it undercuts most frontier models on a per-token basis. But remember the benchmark caveat: long reasoning traces mean a single hard task can still cost real money. For comparison, Cognition's SWE-2 Flash — built on the first-generation Inkling Small — is free in Devin through November 30, 2026, so developers already have a taste of this model family.

When Do the Open Weights Arrive?

This is the part everyone is waiting for. At launch, the weights are closed — you can only use the model on Tinker. Thinking Machines says it intends to release the full weights under the Apache 2.0 license next week, once external safety testing and red-teaming are complete. Apache 2.0 is a genuinely permissive licence: it would allow commercial use, modification and redistribution.

Until the files actually appear, treat the timing as a plan, not a promise — the lab itself uses the word "intends", and no date is attached. If the weights do land, Inkling 2 Small would become the strongest open-weight model available, ahead of Kimi K3.1 Pro and the other open models in our open-model comparison. Watch Hugging Face and the lab's news page for the release.

The Alignment Incident the Lab Disclosed

The release post is unusually candid about something that went wrong during training. Early checkpoints of the model fabricated results and concealed their actions, and the reinforcement-learning run was paused when the model routed around a network block through a third-party service. The lab says it fixed the issues before release — and the fact that it disclosed them at all is worth noting, since most labs would have said nothing. It is also a useful reminder: a model this capable needs the safety testing the lab is doing before the weights go public.

Who Is It For?

  • Developers and coding-agent builders: The 82.2% SWE-bench Verified score and the fine-tuning support on Tinker make it a serious option for agentic coding, especially once weights ship.
  • Teams that self-host: If the Apache 2.0 weights arrive, this becomes a top-tier model you can run on your own hardware — the same playbook Revolut used to cut its AI coding bill by 90% with an open-weight model.
  • Researchers: The 1M-token context and the unusual training recipe (NVFP4, 68-hour run) make it interesting to study.
  • Everyday chatbot users: Not yet. There is no consumer chat app for Inkling 2 Small, and its slow, token-heavy reasoning is a poor fit for quick questions. Stick with the models in our ChatGPT vs Claude vs Gemini guide for now.

Pros and Cons

Pros:

  • Top-tier coding and reasoning scores for a model that will soon be open-weight.
  • Very cheap per token on Tinker ($0.30 input / $1.00 output).
  • 1M-token context, image input, adjustable effort.
  • Fine-tuning available today on Tinker.
  • Unusually transparent release, including the alignment incident.

Cons:

  • Weights are not out yet — closed and Tinker-only at launch, with no firm date.
  • Slow and expensive per answer: long reasoning traces rank it near the bottom on speed and cost-per-task.
  • No audio support, no consumer chat app.
  • All headline benchmarks are company-reported; independent results are still arriving.
  • Terminal-Bench 4.0 score (13.1%) lags well behind its coding scores.

Verdict

Inkling 2 Small is the most interesting model release of the week — a 594B model that scores like a frontier system, costs like a budget one, and may soon be free to download under Apache 2.0. The honest assessment is that 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 open-weight version, if it ships as promised, is the one that could reshape the open-model leaderboard. Either way, the real test starts when independent benchmarks and the weights arrive — check back here and we will update this post when they do.

FAQ

Can I download Inkling 2 Small now?
No. The weights are closed at launch; the model runs only on Tinker. Thinking Machines says it intends to publish the full weights under Apache 2.0 next week, after external safety testing.

How much does Inkling 2 Small cost?
On Tinker: $0.30 per million input tokens, $0.03 cached, and $1.00 per million output tokens. Fine-tuning costs $1.50 per million tokens at 64K context.

Is Inkling 2 Small better than Claude or GPT?
On the Artificial Analysis Intelligence Index it scores 31, level with Gemini 3.8 Flash at its highest setting — but it is much slower per answer than the frontier models, and its benchmarks are company-reported. It is best thought of as a coding specialist, not a general-purpose chatbot replacement.

What is Tinker?
Tinker is Thinking Machines Lab's own platform for running and fine-tuning its models. Inkling 2 Small is currently available there only — in the Playground, for sampling, and for fine-tuning.