Thinking Machines' Inkling 2 Small Is a 594B Model You Can Only Use on Tinker — for Now

Thinking Machines Lab has launched Inkling 2 Small, a 594-billion-parameter model that scores 31 on the Artificial Analysis Intelligence Index — its best result yet — but you cannot download it. The weights are closed at launch and the model is available only through Tinker, the company's own platform. Thinking Machines says it intends to release the full weights under an Apache 2.0 licence next week, after external safety testing and red-teaming are complete.

Inkling 2 Small is the second-generation model from Thinking Machines Lab, the startup founded by former OpenAI chief technology officer Mira Murati. It is a Mixture-of-Experts model, or MoE: of its 594 billion total parameters, only 17 billion are active for any given token, which is how a model this large can run at a practical cost. It supports image inputs and variable effort, letting users trade compute for reasoning depth on harder tasks. The announcement came on 5 November 2026 in US time, 6 November in Pakistan.

The launch is a study in controlled openness. Thinking Machines is promising an open-weights release, but it is not making one today — and the difference between "intends to release next week" and "released" is exactly the kind of distinction this blog checks before publishing. Here is what is confirmed, what is promised, and what the 31-point score actually means.

Illustration created for INORYVA

What the numbers say

The headline figures are straightforward. Inkling 2 Small has 594 billion total parameters with 17 billion active — a ratio that puts it in the same design family as other large MoE models, where most of the model's knowledge sits in experts that are only consulted when relevant. Thinking Machines says the model "far surpasses" its previous models' scores on the Artificial Analysis Intelligence Index, reaching 31, with gains across agentic coding, vision, knowledge work, and reasoning.

The Artificial Analysis Intelligence Index is a composite benchmark score that combines results across many tasks into a single number, and it is one of the most widely cited ways to compare frontier models. A score of 31 puts Inkling 2 Small in serious company, but the number needs its context: Thinking Machines' own comparison table lists the score as 30.6 on the previous version of the index, and 31 on the current one. This article uses 31 only, the current-version figure, and does not present both numbers as if they were two different results.

Specification	Inkling 2 Small
Total parameters	594 billion (MoE)
Active parameters per token	17 billion
Artificial Analysis Intelligence Index	31
Inputs	Text and images
Effort	Variable (user-controlled)
Weights at launch	Closed
Planned licence	Apache 2.0 (next week, intended)
Access at launch	Tinker only
Tinker pricing	$0.30 per 1M input tokens; $1.00 per 1M output tokens

For comparison, the model sits in a busy week for large-model releases. OpenAI's GPT-6 Astra scored 62.7 on the benchmarks OpenAI chose to highlight, and Moonshot's Kimi K3.1 Pro arrived with its own large-scale claims — but those are different benchmarks and different index versions, so direct score-to-score comparisons across announcements are misleading. The honest comparison is the one Thinking Machines makes itself: this model is a large step up from its own previous models, and the index score is the evidence.

Tinker only — and what Tinker is

At launch, Inkling 2 Small is available only on Tinker, Thinking Machines' platform for running and fine-tuning models. You can try it on the Tinker Playground today, and it is also available for fine-tuning and sampling on Tinker. There is no download, no Hugging Face repository, and no weights file to run locally — that is what "closed weights" means in practice.

Tinker pricing for the model is $0.30 per million input tokens and $1.00 per million output tokens. That is the API-style cost of using the model through Thinking Machines' own infrastructure, and it is the only way to run it until the weights release. For developers, the practical question is whether to build on Tinker now or wait for the Apache 2.0 release and run the model themselves. The answer depends on how much the one-week wait matters — and on whether the weights release actually lands next week, since Thinking Machines frames it as an intention, not a guarantee.

The Tinker-only launch also explains why the model matters to the open-model community even before its weights exist. Thinking Machines is signalling that the weights are coming, and that the delay is for safety testing rather than a change of strategy. The company's own words are: "We intend to make the full weights available under an Apache 2.0 license next week after the completion of external safety testing and red-teaming." Intend, not will — and "after the completion" means the schedule depends on the testing, not the calendar.

Illustration created for INORYVA

The Fledge-Alpha clue

One detail in the announcement gives an independent check on the model's quality. Thinking Machines says it tested early checkpoints of Inkling 2 Small on OpenCode under the name "Fledge-Alpha". OpenCode is a popular open-source coding agent, and its public leaderboards have tracked a mystery model called Fledge-Alpha for some time. The disclosure connects that leaderboard presence to Inkling 2 Small, and it means the model's coding ability has already been visible in the wild — just under a different name.

That matters because benchmark scores in a vendor announcement are, by definition, the vendor's own numbers. A model that has been running on a public coding tool under a pseudonym has a track record that does not depend on the announcement. It does not prove the 31-point index score, but it does suggest the agentic coding gains Thinking Machines claims are real enough to have shown up in independent use.

Readers following the model landscape will recognise the pattern from the Carbon vs Opus 5.5 vs Astra flagship comparison, where the headline scores only made sense once the benchmarks behind them were explained. The same discipline applies here: Inkling 2 Small's score is a claim about a composite index, and the Fledge-Alpha detail is the most useful independent signal in the announcement.

What to watch next week

The test of this launch is not today's announcement but next week's weights release. Three things to check when it lands: whether the full weights actually appear under Apache 2.0, which permits commercial use and modification; whether the released weights match the model running on Tinker, including the 594B/17B architecture; and whether independent evaluators can reproduce the index score of 31 on their own hardware. Until those checks are possible, Inkling 2 Small is a closed model with an open promise — impressive on paper, unverified in the open.

Google's Gemini 4 Carbon and the other frontier models in this week's news show how crowded the top of the market has become. Inkling 2 Small's distinctive bet is efficiency: a very large model that only activates a small fraction of itself per token, offered first through the company's own platform and then, if the plan holds, as open weights. If the weights land as promised, it becomes one of the largest open models available — and the Tinker-only week will look like a brief safety pause. If they slip, the announcement will read very differently.

Illustration created for INORYVA

Key Takeaways

• Inkling 2 Small is a Mixture-of-Experts model with 594 billion total parameters and 17 billion active per token, supporting image inputs and variable effort.
• It scores 31 on the Artificial Analysis Intelligence Index, which Thinking Machines says far surpasses its previous models, with gains in agentic coding, vision, knowledge work, and reasoning.
• The weights are CLOSED at launch. The model is available only through Tinker — Playground, fine-tuning, and sampling — at $0.30 per 1M input tokens and $1.00 per 1M output tokens.
• Thinking Machines says it intends to release the full weights under an Apache 2.0 licence next week, after external safety testing and red-teaming — an intention, not a firm date.
• Early checkpoints were tested on OpenCode under the name "Fledge-Alpha", giving the model's coding ability a public track record before the announcement.
• Do not call this an open-weights release yet. The weights are not available, and the release depends on safety testing finishing first.

Conclusion

Inkling 2 Small is a strong model with a careful launch: a 594B/17B MoE design, a 31-point index score, and a Tinker-only debut that keeps the weights closed while safety testing runs. The promise that matters is the Apache 2.0 release planned for next week — if it lands, Thinking Machines will have delivered one of the largest open models ever released, and the closed week will look like a responsible pause. If it slips, the announcement's careful wording will be the tell. For now, the model is available to try on Tinker, the score is a vendor-reported composite, and the open-weights claim is a plan, not a fact. Check the weights, not the press release, next week.

Sources

• Thinking Machines Lab — "Introducing Inkling 2 Small" — https://thinkingmachines.ai/news/inkling-2-small/
• Thinking Machines Lab — Tinker pricing — https://thinkingmachines.ai/tinker/pricing/
• The Tribune India — "Thinking Machines unveils Inkling 2 Small model with 594 billion parameters" — https://www.tribuneindia.com/news/business/thinking-machines-unveils-inkling-2-small-model-with-594-billion-parameters/

Facts checked against Thinking Machines' own announcement on 6 November 2026. The launch date (5 Nov US time) is dated via secondary coverage, as the announcement page shows no date line.