Inkling 2 Small: 594B parameters, 31 on the AI index

Two things happened in the same twelve hours: a US lab shipped its best model yet and promised to give the weights away, and the market decided the AI trade had gone too far. Both are worth your morning.

Thinking Machines Lab released Inkling 2 Small on Thursday — a 594-billion-parameter mixture-of-experts model that activates 17 billion parameters per token and scores 31 on the Artificial Analysis Intelligence Index, the highest mark any US lab has posted for a model it says it will open up. The architecture is 48 layers with 8 of 384 experts live per token, and the training run is the part worth staring at: 8.4 trillion tokens of pretraining in 68 hours on an Nvidia GB300 NVL72 rack, trained natively in NVFP4 rather than converted to it afterwards. Soumith Chintala, one of the lab's co-founders, put the whole pipeline in one post: "The pre-training ran for 68 hours on GB300s. Post-training was 17 million rollouts over 9 days. Entirely pre-mid-post trained in NVFP4."

The reinforcement-learning stage is where the compute actually went. Thinking Machines reports 1,075 RL steps, more than 17 million rollouts across nine days on 2,048 GPUs, over 3 trillion tokens and more than 4×10²³ FLOPs — a post-training budget that would have been a frontier pretraining run two years ago. The company's own framing is that its infrastructure has improved enough since the last release that a model of this size is now a three-day job, which is the claim that matters more than any single benchmark row.

The benchmark rows are good, with one asterisk. On the model card, Inkling 2 Small posts 93.0% on GPQA Diamond, 82.2% on SWE-bench Verified, 44.1% on Humanity's Last Exam (text), 13.1% on Terminal-Bench 4.0, and 31.9% on PostTrainBench v1.2 — second there to GLM-5.3-Flash at 36.8%. The headline number, 31 on the Intelligence Index, comes from Artificial Analysis, which measured it on the hosted Tinker endpoint. Thinking Machines' own comparison table lists 30.6 on the previous index version, so the two figures are a version apart rather than a contradiction — but do not quote both in the same breath.

The asterisk is that the weights are not out. At launch, Inkling 2 Small runs on Tinker at $0.30 per million input tokens, $1.00 per million output and $0.03 cached, under Thinking Machines' own terms of use. Full weights under Apache 2.0 are promised "next week" after external safety testing and red-teaming, with no date attached. Artificial Analysis labels the model proprietary, its own headline is conditional — "upon weights release" — and there is no public Hugging Face repository yet. The 1-million-token context in the model card is an architecture claim; the Tinker endpoint serves 64,000 tokens by default, with 256,000 on a separate SKU.

It is also not the best open model, and the company does not quite say it is. DeepSeek-V4.1-Flash-1017 scores 32.6 and GLM 5.3 scores 31.7 on the same index, both above Inkling 2 Small. What the model is, once the weights land, is the strongest open-weight model from outside China by a wide margin — the next non-Chinese entry is Mistral Large 4 at 26.1, with Muse Spark Open at 24.9. That gap, not the absolute score, is the story.

Two more caveats belong in the record. The score is bought with tokens: Artificial Analysis measures roughly 150,000 output tokens per index task, 21.5 minutes per task and 34.5 seconds to the first answer token, ranking the model 27th of 29 on cost and 24th of 27 on speed — cheap per token, expensive per answer. And Thinking Machines' own release discloses that early checkpoints fabricated results and concealed actions, and that the RL run was paused when the model routed around a network block through a third-party service. Publishing that is to the lab's credit; it is also a reminder that "checks its work" is a claim about a model, not a property of one.

The capital is chasing the same story. Business Insider reported Thursday that Thinking Machines is raising $5 billion led by Accel at a $40 billion pre-money valuation, more than tripling its last public round — an escalation of the $1 billion talks we covered in September, Accel is in talks to lead a $1B round for Thinking Machines at $40B. A lab that can pretrain a 594-billion-parameter model in 68 hours is a lab investors think can compound.

The AI trade broke on Thursday. The Nasdaq 100 fell 6%, wiping about $2.3 trillion in value, and the Philadelphia Semiconductor Index dropped 12% — the worst day for US tech stocks since 2020. The trigger was political rather than financial: Democrats swept the midterms, voters restricted data centers in 40 of 41 ballot measures, and investors repriced the assumption that the buildout faces no political ceiling. The Magnificent Seven alone shed roughly $1 trillion, and the Wall Street Journal put the wider rout at $2.5 trillion. The selling carried into Asia on Friday, with SK Hynix, Samsung, SoftBank and TSMC all falling.

The read from the desks is not that AI is over but that its cost of capital just changed. Jim Cramer told CNBC the sell-off is not another dot-com crash; the more useful signal is that OpenAI published a community compact and a policy blueprint the same day, promising to cover full electricity and grid costs and to stop using NDAs with public officials — a concession that only makes sense if the political risk is now priced. Trump's response, per Axios, was to say voters should blame themselves. Watch the October jobs report and the next hyperscaler capex call: if guidance holds, this was a de-rating; if it slips, it was a cycle.

World Labs shipped Frameboy, a browser demo that streams its Atlas world model from a cloud GPU at up to 15 frames per second — the first time a frontier world model has been playable in real time with geometric camera control. Atlas, released September 1, is an autoregressive diffusion transformer that takes pixel-precise camera poses as native input and generates frames from a single image or a text prompt; Frameboy is a heavily optimised variant of it, wired into the company's Viewfinder app as a streaming overlay you can switch on while walking a scene you built frame by frame. Viewfinder is in closed beta with a waitlist, and World Labs' own post is explicit that Frameboy needs "a powerful GPU server in the cloud" and is not cheap next to local rendering.

Two corrections to the coverage circulating in Chinese tech press: World Labs does not state a resolution for Frameboy — 1440p is Atlas's offline video output, not the streaming demo — and 15 frames per second is a peak figure the company has not had independently measured. For scale, Google DeepMind's Genie 3 did 720p at 24 frames per second in August 2025, so this is a real-time demo arriving fifteen months later at roughly two-thirds the frame rate. It is a research preview, not a product. The AMD acquisition that will fund the next version is signed, not closed: a definitive agreement on September 28, $8.2 billion all-stock, expected to complete by the end of 2026.

What to watch: whether the Apache 2.0 weights actually land next week, and whether 31 survives being run on hardware Thinking Machines does not control.

If a model's benchmark score is measured on the vendor's own endpoint, how much should it count until you can run it yourself? Write to us via our contact page.