Who is Thinking Machines Lab
Thinking Machines Lab was founded in February 2025 by Mira Murati, the former chief technology officer of OpenAI, alongside fellow OpenAI alumni John Schulman, Barrett Zoph, Lilian Weng, Andrew Tulloch, and Luke Metz. Murati had spent years at OpenAI overseeing the development and shipping of ChatGPT, DALL-E, and GPT-4 before leaving in 2024 and going on to found her own lab.
Within five months of founding, Thinking Machines closed one of the largest seed rounds in venture-capital history — around $2B at a $12B valuation, led by Andreessen Horowitz with participation from Nvidia, AMD, Cisco, and Jane Street. Rather than chasing the biggest model on the leaderboard, the lab has publicly framed its bet as building AI systems that are more customizable, more transparent about their reasoning, and cheaper to adapt — a deliberate contrast to the "one bigger model to rule them all" strategy of the frontier labs.
Tinker, then Inkling
Their Tinker platform is a production fine-tuning service — it lets developers fine-tune models without worrying about the cost or complexity of distributed training infrastructure. Inkling is their first foundation model — and they trained it from scratch, not as a fine-tune of someone else's weights.
The numbers that matter
975B total / 41B active MoE. 45 trillion tokens of pretraining — text, images, audio, and video natively. 1M token context. 30M+ RL rollouts. NVIDIA GB300 NVL72 hardware. Apache-2.0 on Hugging Face.
Honest about the leaderboard
The benchmark table is unusually comprehensive and unusually honest about where Inkling sits. Against open-weight peers it's competitive throughout. Against the closed frontier (Fable 5, GPT-5.6 Sol) it's clearly below. They say this plainly: "It is not the most performant model available today, closed or open." That admission, made in the opening paragraph, is rare.
77.6% SWE-Bench Verified — strong for an open model, matching Kimi K2.5 and GPT-OSS. 54.3% SWE-Bench Pro — mid-pack. 87.9% GPQA Diamond — competitive with the best open models. 97.1% AIME 2026 — near-saturated like everyone else. 46% HLE with tools — respectable, well below closed frontier.
Update — July 30: the small sibling surpasses the big one, on purpose
Inkling launched with a preview of a second model, promising full weights "once testing is complete." Thinking Machines has now shipped it: Inkling-Small, 276B total / 12B active — their own framing is "comparable performance to Inkling at a quarter of its size."
That framing undersells it. Inkling-Small doesn't just match Inkling — on reasoning and agentic coding, it surpasses it: HLE text-only 31.6% against Inkling's 29.7%, GPQA Diamond 89.5% vs 87.2%, SWE-Bench Verified 80.2% vs 77.6%, ARC-AGI-1 84.0% vs 79.5%, ARC-AGI-2 40.1% vs 36.5%. Every one of those is the smaller model ahead of the one nearly 3.5× its size.
The reason isn't mysterious once Thinking Machines explains it, and the explanation is the actual news here. Inkling-Small isn't Inkling-shrunk — it's Inkling-after: it started training once the larger model was already done, benefiting from fixes to the pretraining data mix and the ML recipe. Post-training went further still — an earlier checkpoint, "Inkling-Small (preview)," was partly trained via on-policy distillation with Inkling itself as the teacher, then given two additional weeks of agentic-coding RL on top. So the smaller model inherited the bigger one's judgment and then kept training past it on exactly the tasks where it now leads.
What it gives up is knowledge, not reasoning — and Thinking Machines states this as plainly as they did Inkling's original weaknesses: "Inkling maintains an advantage on knowledge coverage and factuality." SimpleQA Verified drops to 20.6% from Inkling's 43.9%; AA Omniscience goes from a positive 2.1 to −9.0. That's the coherent story: distillation and RL transfer reasoning process, but raw factual recall tracks parameters memorized during pretraining, which a smaller model has less room for. Reasoning is a skill you can hand down; trivia is not.
One thing the announcement doesn't dwell on is worth naming, since the whole point of these updates is not letting a good headline hide the rest of the table. FORTRESS adversarial — whether the model refuses genuinely harmful requests — drops to 71.6% from Inkling's 78.0%, described in the post only as "competitive." A 6.4-point fall in refusing harmful requests is a real number worth a reader's attention on its own, not folded into a one-word characterization.
For workloads it was built for — high-volume coding, using an LLM as a grader, generating synthetic data — this is close to a free upgrade over the flagship at a quarter of the compute. For anything that leans on broad world knowledge, Inkling still leads.
Update — July 31: the framework behind that "competitive" FORTRESS number
The 6.4-point drop in refusing harmful requests flagged above, folded into one word in the original announcement, now has its full reasoning spelled out. Thinking Machines published the safety framework it used to decide Inkling's weights were safe to release: internal evals across CBRN/cyber/misuse/multimodal harms, and four external red-teaming firms each assigned a different risk lane. The methodologically important part: adversarial fine-tuning that strips out refusal behavior entirely, to test whether the model reveals new dangerous capability once its safeguards are gone. It didn't. That's the actual claim behind "competitive": not that Inkling is risk-free, but that removing its refusals doesn't unlock anything existing open-weight models don't already offer.