Thinking Machines Lab ships Inkling 2 Small, a 594B MoE with 17B active parameters
On November 5, 2026, Thinking Machines Lab introduced Inkling 2 Small, a mixture-of-experts model with 594 billion total parameters and 17 billion active, with image input and variable effort, first available under closed weights on Tinker. For teams evaluating an open-weight ‘flash’ model for agentic coding, the real story is not the parameter total but the gap between today’s closed access and the Apache 2.0 weights promised next week.
November 5, 2026. Thinking Machines Lab introduces Inkling 2 Small, its second model generation. 594 billion. The total parameter count of the mixture-of-experts. 17 billion. The active parameters per inference. 31. Its score on the Artificial Analysis index, above every previous model from the lab. Why it matters: the word “Small” hides the real story — the open-weight race now turns on active parameters, not totals, and on the weights-release strategy.
594 billion parameters, 17 billion active: why the name misleads
Inkling 2 Small is a Mixture-of-Experts with 594 billion total parameters and 17 billion active per pass. That is the architecture that has become standard in the “flash” tier: a large reservoir of capability, but an inference cost bounded by the number of experts consulted per token. The model accepts image inputs and variable effort — the user can ask for more reasoning, or less, depending on the cost they are willing to pay.
The “Small” label is a naming legacy, not a description. Next to GLM-5.3-Flash, DeepSeek-V4.1-Flash, or Kimi K3, this model is compared on active parameters and cost per token, not on total parameters, which no longer has much to do with the real price of a query.
The lab claims a score of 31 on the Artificial Analysis index, “far surpassing” its previous models, with gains in agentic coding, vision, knowledge work, and reasoning. The improvement comes, according to the announcement, from more efficient pre-training and a large, unified reinforcement-learning run.
What it actually does: from agentic coding to autoresearch
The published demos are not screenshots: they are complete artifacts the model generated. Inkling 2 Small builds a city simulator in Three.js with no pre-made assets — every building, texture, and car is generated procedurally, with power, budget, and emergency management. It writes a PC emulator in Python that boots a real Linux kernel in the browser. It generates an origami scene in Blender, or a 3D floor plan from a map.
The core skill remains agentic coding. The model was trained with RL on hundreds of thousands of long-horizon coding tasks, using an agentic grader that assesses the approach, not just unit-test pass rates. On the reference benchmarks, it lands consistently second: 25.4% on FrontierCode v2.0, 72.1% on DeepSWE v1.1, 86.3% on SWE-bench Pro v2 (Hard) — each time behind a different competitor, never the same one.
The most revealing part is autoresearch. On PMPP-Hard, a GPU kernel-optimization benchmark, Inkling 2 Small solves an average of 54.25 of 69 tasks in ten hours — behind GPT-6 Astra (63), Kimi K3.1 Pro (60.5), and Claude Opus 5.5 (60), but level with DeepSeek V4.1 Flash (55.25). The lab also shows a case study: the model autonomously trained a Qwen3.5-4B on Tinker to forecast time series, in 4.5 hours for $426, doubling the starting forecasting score.
The positioning: the open-weight “flash” race
The comparison table published by Thinking Machines Lab explicitly places Inkling 2 Small in a category that Chinese labs dominate: open-weight “flash” models. Against GLM-5.3-Flash, DeepSeek-V4.1-Flash-1017, Muse Spark Open, and Kimi K3, the Thinking Machines model is rarely first, but rarely far behind.
That positioning is deliberate and coherent. The open-weight “flash” tier rewards cost-to-capability, not the absolute score. A model with 17 billion active parameters that holds second place on most coding benchmarks, with native image input, is exactly the profile a team can self-host on one or two cards for its internal agent loops.
With one caveat, and it is a big one: the weights are not open yet.
The weights timing: closed today, Apache 2.0 next week
Inkling 2 Small follows a two-step release sequence. Today it is only accessible through Tinker, the lab’s platform, under closed weights. Thinking Machines Lab says it intends to release the full weights under an Apache 2.0 license next week, after “external safety testing and red-teaming.”
That window is the single most important data point for a decision-maker. It materializes the lab’s stance, formalized in its post “A safe path to open weights”: open the weights after external validation, not before. Fine-tuning and sampling are already available on Tinker, and checkpoints will be downloadable with the weights release.
The risk for an adopter is the gap between evaluation and availability. You can test the model today, measure its behavior on your own tasks, and plan self-hosting — but you cannot run it locally until the weights ship, on a non-binding date.
The acknowledged limits
The announcement lists the flaws without hedging. Inkling 2 Small is “not the most succinct communicator” and does not always choose the right tool, especially in codebases that expose many skills. Like previous Inkling models, it accepts images natively but does not support audio. It is also “less optimized for the chat experience,” with organization issues, overuse of slashes, and overreaction to some safety-related queries.
Those limits draw a usage profile: an execution and coding model, not a chat assistant. That is consistent with the open-weight “flash” positioning, where the model is a component in an agent pipeline, not an interface.
The context: an American lab in a China-dominated tier
The arrival of Inkling 2 Small matters beyond the model itself. Thinking Machines Lab, founded by Mira Murati after leaving OpenAI, is one of the few American labs fighting in the open-weight “flash” tier — ground that GLM, DeepSeek, Qwen, and MiniMax hold in force. Most American players concentrate their open models at the very large or the very small end; few ship a mid-size MoE at 17 billion active parameters with native image input.
That presence matters, because the open-weight “flash” tier is exactly the one enterprises can self-host at reasonable cost. When an American lab enters it with a promised Apache 2.0 license, it offers a sovereignty alternative to teams that did not want to depend on a Chinese vendor — a commercial argument as much as a technical one. Whether Thinking Machines Lab can keep the release cadence against labs shipping a model every quarter remains to be seen.
Verdict
Inkling 2 Small is a credible entry in the open-weight “flash” tier, with a profile that is rare from an American lab: a 594-billion total / 17-billion active MoE, native image input, and a consistent second place on agentic-coding benchmarks.
If you already evaluate GLM-5.3-Flash or DeepSeek-V4.1-Flash for self-hosted agentic coding, test Inkling 2 Small on Tinker now — its scores and cost put it in the same pack, and native image input may tip the balance on your visual cases. If you need to deploy it locally, wait for the Apache 2.0 weights next week: evaluating today is useful, deploying is impossible before. If your criterion is absolute score, look at the “pro” models instead — Kimi K3 or Claude Opus 5.5 remain ahead on the hardest tasks. The rule to remember: in the flash tier, the number that counts is no longer the parameter total, but the active count — and the license that will come with it.