Thinking Machines Inkling 2 Small: 594B model, Apache 2.0 weights promised next week, benchmarks self-reported

Thinking Machines Lab released Inkling 2 Small on November 5, 2026. It is a large mixture-of-experts model with a promise attached: full weights under an Apache 2.0 license "next week," after safety testing. If the promise holds, it would be one of the most capable openly licensed models available. Every score below is the lab's own, and none has yet been reproduced independently.

The specifications

The lab says Inkling 2 Small has 594 billion total parameters, with 17 billion active for each token. It accepts image inputs and does not support audio, and it offers variable effort settings. The vision encoder has 292 million parameters and is based on Perception Encoder.

The training details are unusually specific. Pretraining used 8.4 trillion tokens and ran for 68 hours on Nvidia GB300 NVL72 racks, natively in NVFP4, a four-bit format. Later phases used 512K and 1M context lengths. The main reinforcement learning run used 1,075 steps, more than 17 million rollouts, 2,048 GPUs and nine days, with 256K context.

The scores

The lab reports these figures, which we list as claims:

• Artificial Analysis Intelligence Index: 31 (the benchmark table lists 30.6).
• SWE-Bench Pro v2 (Hard): 86.3%.
• DeepSWE v1.1: 72.1%.
• FrontierCode v2.0: 25.4%.
• MMMU Pro: 81.6%.
• HLE-Diamond: 28.7%.
• PMPP-Hard kernel optimization (AUC): 494.27, second highest in its table.
• DesignArena Website: 1340, third on the leaderboard snapshot.

Context matters for the index figure. The Artificial Analysis index is the number we reference for comparing models, and a score of 31 puts Inkling 2 Small well below the closed flagships that the lab itself names as ahead of it. Our explainer on what AI leaderboards rank and what a top score hides covers how to read them; the key point is that a self-reported entry is a claim until the leaderboard operator runs it.

The licence is the story

Closed for now, open in a week: that is the commitment. Thinking Machines says it will "make the full weights available under an Apache 2.0 license next week." Apache 2.0 is a permissive license that allows commercial use and modification, which makes it more open than the custom licenses attached to several recent large releases. Our explainer on open weights versus open source covers why the license text, not the label, decides what you can do.

For now the model is available on the company's Tinker platform for testing, sampling and fine-tuning. The announcement does not give a price, though it links to Tinker's pricing page. It mentions one case study with $426 of Tinker spend for a single autoresearch task.

Safety testing

The lab evaluated an internal helpful-only variant next to the release checkpoint and says it used more external testers than for any previous release. It reports a FORTRESS average risk score of 16.2 and an overrefusal score of 20.4, where lower is better for both. Its internal misbehavior rate is 14.9, versus 66.8 for the earlier Inkling model. On cyber, it reports 72.7% on Vals CyberBench and 79.6% on PatchEval.

The lab is candid about a limit. It wrote that "Without proper attention to alignment, capabilities-focused RL can push models to succeed at any cost," and that Inkling 2 Small trails closed models on its alignment evaluations. That candor is useful. It is also a reason to wait: an open-weight release cannot be recalled if a flaw appears later.

How it fits the open-weight race

The open-weight field is crowded and fast. In the past week alone, our desk logged Mistral's trillion-parameter release and Moonshot's Kimi K3.1; see our pieces on Mistral Large 4 and Kimi K3.1 Pro. Thinking Machines is the only one of the three that is a US lab promising a fully permissive license, and that is its distinguishing claim.

A "Small" model at an index score of 31 is not at the frontier, and the lab does not claim it is. What it signals is where open models sit relative to closed ones, and how fast the gap closes.

What the training run tells you

The training numbers are worth a second look because they show how cheap the post-training stage has become relative to the headline size. A nine-day reinforcement learning run on 2,048 GPUs is within reach of a well-funded startup, not only of the largest labs. The 68-hour pretraining run in a four-bit format suggests the lab is betting on efficiency of hardware use over sheer scale.

Neither number establishes quality. They establish that a model of this size can be produced on the schedule of a product cycle, which is the real pressure on closed labs: if a competitor can ship a capable open model every few months, the premium on a closed one has to be justified by something other than access.

What to check next week

Three things settle the question: that the weights appear on the promised date, that the license file reads Apache 2.0 without added use restrictions, and that an independent group reproduces the headline coding scores.

Our take

A useful release with a license promise worth holding the lab to. We would not credit the benchmark numbers until Artificial Analysis or another independent party runs them. We would also watch the one-week clock: if the weights arrive on schedule under the stated license, with the safety report intact, trust in the lab goes up. If the date slips, so does the case for treating this as an open model at all.