Thinking Machines Lab
Thinking Machines Lab is the San Francisco Bay Area AI lab founded in February 2025 by former OpenAI CTO Mira Murati. It raised the largest seed round on record — $2 billion at a $12 billion valuation, led by Andreessen Horowitz — and ships open-weight, natively multimodal models: Inkling (July 2026) and Inkling-Small, both Apache 2.0, alongside Tinker, its fine-tuning service.
Thinking Machines Lab is the AI lab Mira Murati — OpenAI’s former CTO — founded in February 2025, and the newest US entrant to ship real open-weight models. Its founding bet is different from the frontier labs’: rather than racing for the strongest general model, it argues AI should be customised to each user and organisation, and it builds accordingly — open weights you can fine-tune, and Tinker, a service that makes fine-tuning large models practical without your own GPU cluster (TechCrunch).
The money behind it is unlike any startup before it: a $2 billion seed round — the largest on record — closed in July 2025 at a $12 billion post-money valuation, led by Andreessen Horowitz with participation from Nvidia, Accel, ServiceNow, Cisco, AMD and Jane Street (TechCrunch). In July 2026 the bet produced its first models: Inkling and Inkling-Small, open-weight mixture-of-experts models under Apache 2.0 that reason natively over text, images and audio — a combination no other open-weight model in our ranking offers.
Quick facts
| Founded | February 2025 |
| Founder & CEO | Mira Murati (former OpenAI CTO) |
| Chief scientist | John Schulman (OpenAI co-founder) |
| Headquarters | San Francisco Bay Area, California |
| Funding | $2B seed (July 2025) — the largest seed round on record |
| Valuation | $12 billion post-money at the seed |
| Backers | Andreessen Horowitz (lead), Nvidia, Accel, ServiceNow, Cisco, AMD, Jane Street |
| Flagship models | Inkling and Inkling-Small — open weights, Apache 2.0 |
| Key products | Tinker (fine-tuning service), Tinker Playground, the Inkling models |
| Licence model | Apache 2.0 open weights; revenue from the hosted Tinker service |
| Notable | First models shipped ~17 months after founding; team drawn heavily from OpenAI |
History and founding
Murati left OpenAI in September 2024 after serving as CTO through the ChatGPT era (and briefly as interim CEO during the November 2023 board crisis), and founded Thinking Machines Lab in February 2025. The founding team drew heavily on OpenAI: John Schulman, an OpenAI co-founder and a principal architect of RLHF, joined as chief scientist, alongside researchers including Barret Zoph and Luke Metz (Built In).
The company describes its mission as building “collaborative general intelligence” — AI that works with people and adapts to them, rather than a single monolithic assistant. In November 2025, Bloomberg reported the company was in talks to raise at a valuation of around $50 billion (Reuters via Yahoo Finance); no such round has been announced since, and the last confirmed valuation remains the seed’s $12 billion.
Models and products
| Product | What it is | Notes |
|---|---|---|
| Inkling | Open-weight flagship (15 July 2026) | 975B-param MoE (41B active), 1M context, text + image + audio input, Apache 2.0 |
| Inkling-Small | Compact sibling (30 July 2026) | 276B MoE (12B active) — beats its larger parent on several reasoning and agentic-coding benchmarks |
| Tinker | Fine-tuning service (October 2025) | Fine-tune large open models via API without your own GPU infrastructure |
| Tinker Playground | Hosted chat interface | Try the Inkling models with text, image and audio input |
The models are the argument made concrete. Both Inklings ship as Apache 2.0 open weights on Hugging Face, are natively multimodal — text, images and audio in one model, trained on 45 trillion tokens that included video — and expose a variable thinking-effort dial rather than fixed reasoning tiers. Hosted inference is available through partners: Together AI, Fireworks, Modal, Databricks and Baseten. Full specs and benchmarks are on the Inkling model page.
Tinker (October 2025) is the business: a service that streamlines fine-tuning large language models so teams can customise a model without massive GPU infrastructure. The Inkling models are designed to be its raw material — chief scientist John Schulman has said the company plans to release more models through 2026 (Built In).
Business and market position
Thinking Machines is a private company monetising through Tinker rather than a consumer app — there is no ChatGPT-style assistant, and no per-seat subscription. Its competitive position is distinctive: it is the only major US lab whose entire model line is open-weight, multimodal and permissively licensed, which sets it against the Chinese open-weight cluster (DeepSeek, Moonshot, Zhipu, MiniMax) more than against OpenAI or Anthropic. Simon Willison called Inkling “a new viable contender to join NVIDIA Nemotron and Gemma 4” in the US open-weights ecosystem (simonwillison.net).
Where Thinking Machines excels
- Open multimodality. The Inklings are the only open-weight models in our ranking that reason natively over text, images and audio — the Chinese open cluster is text-only apart from Kimi K3’s vision.
- Clean licensing. Apache 2.0, with none of the custom-licence friction of Llama or Kimi K3.
- Efficiency engineering. Inkling-Small matches or beats a model 3.5x its size from the same lab — 80.2% vs 77.6% on SWE-bench Verified (vendor) — which is the clearest public demonstration of its post-training craft.
- The fine-tuning thesis. Tinker plus open weights is a coherent product story: the models exist to be customised, not just consumed.
Where Thinking Machines falls short
- Not the frontier — by its own admission. The company itself states “Inkling is not the strongest overall model available today, open or closed” (simonwillison.net). Independent scoring agrees: Artificial Analysis puts Inkling at 42 on its Intelligence Index — well above the open-weight median of 27, but far below Kimi K3’s 60.
- Thin documentation. Willison’s main criticism: a sparse model card with minimal transparency about training-data sourcing.
- A young, narrow line. Two models, one product, and no SWE-bench Pro or arena history yet — the track record is seventeen months old.
Developer resources
Weights are on Hugging Face under thinkingmachines; hosted APIs run through Together AI, Fireworks, Modal, Databricks and Baseten; fine-tuning is through Tinker, with the Tinker Playground for interactive use. There is no first-party consumer app.
Frequently asked questions
What is Thinking Machines Lab?
An AI lab founded in February 2025 by former OpenAI CTO Mira Murati, built around the thesis that AI should be customised rather than one-size-fits-all. It ships the open-weight, multimodal Inkling models under Apache 2.0 and sells Tinker, a fine-tuning service.
What models has Thinking Machines released?
Two, both in July 2026: Inkling (15 July, 975B parameters with 41B active) and Inkling-Small (30 July, 276B with 12B active). Both are Apache 2.0 open weights with a 1M-token context that accept text, images and audio — and the smaller model beats the larger one on several reasoning and coding benchmarks.
Is Inkling open source?
The weights are open — released on Hugging Face under the permissive Apache 2.0 licence, free to download, self-host, fine-tune and use commercially. As with most “open” models, the training data and full training code are not published.
How is Thinking Machines funded?
A record $2 billion seed round, closed July 2025 at a $12 billion post-money valuation, led by Andreessen Horowitz with Nvidia, Accel, ServiceNow, Cisco, AMD and Jane Street participating. Bloomberg reported talks at a ~$50 billion valuation in November 2025; no such round has been announced.
How good are the Inkling models?
Mid-pack among open-weight models on independent scoring — Artificial Analysis rates Inkling 42 on its Intelligence Index, well above the open-weight median (27) but well below the open leaders — with vendor-reported SWE-bench Verified scores of 77.6% (Inkling) and 80.2% (Inkling-Small) that sit in the top open cluster. Their distinctive strength is native multimodality: no other open-weight model in our ranking takes text, images and audio.
What is Tinker?
Thinking Machines’ fine-tuning service, launched October 2025 — an API for fine-tuning large open models without running your own GPU infrastructure. It is the company’s main product; the open Inkling models are designed to be customised through it.
Last verified 19 August 2026. Funding, valuation and backers are from TechCrunch’s reporting of the July 2025 seed close; model specs are from Thinking Machines’ own model card and announcements; the independent Intelligence Index figure is Artificial Analysis’s. The reported ~$50B raise is attributed to Bloomberg and has not been announced as closed.
Models
| Model | SWE | Context | In | Out | Status |
|---|---|---|---|---|---|
| Inkling | 77.6% | 1M | $1 | $4.05 | Available |