Thinking Machines Shipped a 594B Model and Punted the Weights to Next Week

Inkling 2 Small launched on November 5. It is a sparse mixture of experts with 594 billion total parameters and 17 billion active per token, a million-token context window, and image input. It scored 86.3 percent on SWE-Bench Pro v2 Hard. It is available through Thinking Machines' Tinker service for inference and fine-tuning.
The weights are not out. The lab says Apache 2.0 weights will follow the week after launch, once external safety testing and red-teaming are done, and it has not given a date.
That gap is the story. Everything interesting about this release lives in the space between what shipped and what was promised.
The model itself
Five hundred and ninety-four billion parameters with 17 billion active puts Inkling 2 Small in the same architectural family as most 2026 frontier releases: you pay for a large pool of knowledge and only compute a slice of it per token. The million-token context and image input put it in the multimodal agent category rather than the chatbot category.
The training run is documented enough to be informative. Thinking Machines trained on GB300 NVL72 racks, completed the initial pretraining phase in 68 hours, then ran a nine-day reinforcement learning workflow with intermediate steps in between. The coding capability came from hundreds of thousands of long-horizon challenges, including a set of what the company calls autoresearch exercises: tasks that required the model to post-train neural networks, optimise kernels, and do other pieces of machine learning work.
The demonstration it chose for those skills is the one I would have picked too. It asked the model to fine-tune Qwen3.5-4B to improve forecasting, and the model doubled the base model's forecasting score in four and a half hours.
That is a small result with a large implication. A model that can improve another model's performance on a specific task, unsupervised, for a few hours, is a model that can be pointed at a repository and told to make it better. Whether it does that reliably is a different question, and the benchmark table is where you would look.
The part almost nobody mentioned
Inkling 2 Small includes a vision encoder, and it accounts for 292 million of the 594 billion parameters. That is a rounding error in the parameter count and a significant capability in practice: the model can look at what it produced.
The example the lab gives is building an e-commerce store. The model writes the site, then checks the rendered product pages to see whether they display correctly, instead of only reading its own code for bugs. That closes a loop that coding models usually leave open. Most agents can tell you the build passed. Fewer can tell you the page looks wrong.
The same loop shows up in image generation, where the model can visually verify the quality of graphics it produces. Vision as a verification channel, rather than as an input modality, is an underrated design choice, and it is the kind of thing that matters more as agents take on longer tasks where the failure is visible but not textual.
Benchmarks, and the claim attached to them
The published numbers put Inkling 2 Small ahead of Spark Open, Kimi K3 and several other open models on three knowledge-work benchmarks, and ahead on several financial task datasets. Artificial Analysis says it matches Gemini 3.8 Flash at high reasoning effort on its index, and that once the weights ship it would be the strongest open model outside China.
That second clause is doing a lot of work, and it is a claim about a model you cannot download yet.
There is also a timing subtlety. On November 4, Ricursive released Ripple-65B Preview, an Apache 2.0 dense model that took the top spot among US open-weight models on Terminal-Bench 2.1 at 80.52. Inkling 2 Small's card reportedly puts it at 81.3 on the same benchmark with the same Terminus harness. So the record moved twice in two days, and only one of the two records comes with downloadable files.
Why holding the weights is defensible
The stated reason is external safety testing and red-teaming before release. That is a legitimate practice. It is also, in this market, a hedge.
Releasing weights is irreversible. A lab that publishes them cannot take them back if a capability turns out to be more dangerous in combination with something else. Waiting a week costs almost nothing commercially, because the API is live and the revenue starts immediately. It gives the lab a public safety narrative and a chance to walk the date back if something goes wrong.
The cost of waiting is that the claim stays unverifiable. "Top open model outside China" is a marketing sentence until there is a torrent. Every week of delay turns a product launch into a promise, and promises get discounted by exactly the amount of uncertainty they carry.
The lab behind it
The commercial context is loud. Business Insider reported that Thinking Machines is raising $5 billion at a $40 billion pre-money valuation, with Accel in talks to lead, which would more than triple the valuation from its last public round. The company raised a $2 billion seed at more than $10 billion last year, and an earlier attempt to raise at $50 billion reportedly ended without a deal.
About a third of the founding team has left, including Barret Zoph, Luke Metz and Andrew Tulloch. That is a lot of departure for a company this young, and it sits oddly next to a $40 billion valuation. The counter-argument is that a research lab is not a band, and what matters is whether the training recipe transfers to the people who stayed.
The timing of the model release and the funding news is not a coincidence. A record on an open-weight leaderboard, a documented training recipe, and a capability demo involving autonomous model improvement is exactly the package you show investors who are being asked to price a lab at 40 times its previous round.
The wider frame
The same week, Ramp's payment data showed enterprise AI token volume up 37.1 percent while spending rose 8.9 percent, with about 1 percent of firms accounting for 80 percent of spend and that top cohort slipping below its summer peak.
A lab selling a 594B model into that market has to answer a hard question: who pays? The coding agent customer is the one with the budget today. That explains why every model release in this cycle leads with SWE-bench and Terminal-Bench and knowledge work, and why the fine-tuning story matters as much as the model card. If your model can make a customer's smaller model better in four hours, you are selling a service, not a checkpoint.
Inkling 2 Small looks like a strong model. The useful test arrives when the tarball does. Until then, the honest summary is that Thinking Machines published a very good set of results, a training recipe worth reading, one genuinely underrated capability, and a date it has not committed to.
Related articles
Chinese text in video always turns to mush. Dreamina handed the typesetting to code
Dreamina launched Motion Studio on November 10. It no longer generates frames; the model writes code that lays out text, shapes and pacing. Chinese titles stop blurring, but whether the content is right is still on you.
A community spreadsheet of taps that turns off the ads inside Chinese Android builds
A no-root tool that switches off the ads Chinese phone makers ship inside the system UI. Its method is a hand-written tap script per phone model, maintained by volunteers every time a ROM moves. It works, and the maintenance is the point.
GPT Image 2.6 held the token price and raised the bill
GPT Image 2.6 kept every token rate from 2.5 and still costs more per image, because more of the work now happens in output tokens. Eleven days into the ramp, the pricing structure says more than the leaderboard does.
Someone rebuilt IDM from the outside and wrote down what they did not touch
A Rust reimplementation of Internet Download Manager claims 5x the speed and 86 languages. The more interesting document is the paragraph explaining how it was built without copying anything.