Thinking Machines
NativeAI NativePrivateUpdated November 15, 2026Mira Murati's post-OpenAI lab — with the highest talent density per dollar in the industry.
- Website
- thinkingmachines.ai
- Founded
- 2025
- CEO
- Mira Murati
- HQ
- San Francisco, CA, United States
- Stage
- Seed
- Revenue
- Not disclosed
- Total Funding
- $2.0B
- Employees
- 501-1000
Executive Summary
Updated Nov 15, 2026
What they do
- Tinker, a managed API for fine-tuning and RL post-training of open-weight models
- Builds its own Inkling open-weight mixture-of-experts models under Apache 2.0 licensing
- Researches real-time 'interaction models' and continual learning for adaptable, customizable AI
Why it matters
- Lets teams own custom models without running distributed GPU training infrastructure
- NVIDIA gigawatt-scale Vera Rubin partnership secures long-term frontier training compute
- Led by ex-OpenAI CTO Mira Murati and chief scientist John Schulman
Best fit
- ML teams with proprietary data needing task-specific models that beat general LLMs
- Enterprises seeking open-weight models they can customize, host and control
- Research groups running RL, reasoning or agent training experiments at scale
Watch-outs
- Several co-founders, including former CTO Barret Zoph, left within the first year
- Inkling models trail top closed frontier models on general intelligence benchmarks
- Enterprise track record is short, with few named large-company production references
Products & Offerings
Tinker
Managed training API for fine-tuning and reinforcement-learning post-training of open-weight models such as Inkling, Qwen, Nemotron, DeepSeek, Kimi and gpt-oss; Tinker runs the GPU infrastructure while customers keep control of their algorithms and data.
Inkling
Open-weight mixture-of-experts language model with 975B parameters and a 1M-token context window, released under Apache 2.0 for enterprise customization.
Inkling 2 Small
Second-generation mixture-of-experts model with 594B total and 17B active parameters, image input and variable effort, aimed at coding and knowledge work; available through Tinker, with an Apache 2.0 weights release planned after external safety testing.
Inkling-Small
Smaller, lower-cost member of the first Inkling open-weight family, built for efficient fine-tuning and deployment.
Tinker Playground
Browser-based environment for sampling and testing Inkling and fine-tuned models before building with the Tinker API.
Recent Releases
- Inkling 2 SmallModelNewsNov 2026
- Inkling 2ModelHugging FaceNov 2026
- InklingModelHugging FaceJul 2026
- MetaModelHugging FaceDec 2025
Latest News
Snowflake · 3 days ago
PositiveCrowdfund Insider · 4 days ago
PositiveHPCwire · 4 days ago
Positivejawlah.co · 8 days ago
PositiveTOKENPOST · 9 days ago
NeutralUpdated Nov 15, 2026
Thinking Machines Lab is a credible, well-funded vendor for teams that want to post-train open-weight models on their own data. Tinker has been in production for a year, offers enterprise encryption and private VPC connectivity, and the company says it does not train on customer content. The Inkling family gives buyers a permissively licensed base model to build on. Run a structured pilot against alternatives such as Together AI, Fireworks AI, OpenAI fine-tuning and Mistral, measuring task-level gains over frontier APIs. Given the company's youth and leadership turnover, avoid making it a sole-source dependency for now.
Signal Provenance
Latest observation date and source for each signal category
| Signal Category | Last Updated | Source |
|---|---|---|
Funding Signals Last funding round, Funding stage | Nov 13, 2026 | Cited web research |
Hiring Velocity Headcount, Headcount growth | Nov 1, 2026 | Research agents (estimate) |
Technical Output Research papers, GitHub activity | Nov 15, 2026 | Hugging Face Hub + news feed + research agents |
News Sentiment News mentions, News sentiment | Nov 15, 2026 | Verified news feed |
Partnership Signals Enterprise partnerships, Distribution partners | Nov 15, 2026 | Research agents (estimate) |
Revenue Indicators ARR estimate, Revenue growth | Nov 15, 2026 | Research agents (estimate) |
Leadership Signals CEO tenure, Founder-led | Nov 15, 2026 | Research agents (estimate) |
Competitive Position Category rank, Market share | Never | Research agents (estimate) |
Why this Bluescore?
6 dimensions, each scored 0–10 against all 250 companies. Each adds its weight × score to the Bluescore — see methodology
Thinking Machines's Bluescore of 56.6 rests most on Investment Momentum (7.3, top 14%) and Product Depth (6.5, top 19%). Its weakest area is Business Model & Revenue (4.4, 45th percentile).
In plain English
Thinking Machines is held back most by Business Model & Revenue, which scores 4.4 of 10 and sits at the 45th percentile among the AI250, with 11.8 left of its 21 points. Its strongest support comes from Investment Momentum at 7.3 of 10 (86th percentile) and Product Depth at 6.5 of 10 (81st percentile). Because Business Model & Revenue carries a 21% weight, lifting that weak score would move its Bluescore more than any other dimension.
Written by AI from the scores below and checked against them; updated nightly.
Traction Signals
Last Round
Seed
Risks
Updated Nov 15, 2026