Inkling 2 Small — Mira Murati's 594B model jumps from 13 to 31 on the AI index, weights after safety tests
Thinking Machines Lab, founded by former OpenAI CTO Mira Murati, introduced Inkling 2 Small on November 5, 2026, the first model of its second generation. It is a sparse mixture-of-experts model with 594 billion total and 17 billion active parameters, takes text and images, and supports adjustable reasoning. It scores 31 on the Artificial Analysis Intelligence Index, up from 13 for its predecessor and ahead of Mistral Large 4 (26), but behind DeepSeek V4.1 Flash and GLM-5.3 (32) and Kimi K3.1 Pro (46). It is closed-weight at launch and available only on Tinker; the company intends to release Apache 2.0 weights in the coming weeks, pending additional safety testing.
The three lines
- Model — 594B-parameter MoE (17B active), text and image input, adjustable reasoning
- Score — 31 on the AA index, up from 13; ahead of Mistral Large 4 (26)
- Catch — closed weights at launch; Apache 2.0 release 'in the coming weeks' after safety tests
Key questions
- inkling 2 small benchmark
- **31 on the Artificial Analysis Intelligence Index — near the top of Western open models, still behind Chinese leaders and closed frontier models.** | Model | AA index | Weights | |---|---|---| | Claude Opus 5.5 | 58 | Closed | | GPT-6 Astra | 57 | Closed | | Kimi K3.1 Pro (China) | 46 | Open | | DeepSeek V4.1 Flash, GLM-5.3 | 32 each | Open | | **Inkling 2 Small** | **31** | Promised | | Mistral Large 4 | 26 | Open | | Meta Muse Spark Open | 25 | Open | | Inkling Small (predecessor) | 13 | — |
- inkling 2 small open weights release date
- **No date yet. The model is closed-weight at launch; weights are promised under Apache 2.0 after additional safety testing.** | Item | Status | |---|---| | Weights at launch | Closed | | Access | Tinker and Tinker Playground | | Planned license | Apache 2.0 | | Timing | 'Coming weeks' (company); some reports say 'next week' | | Condition | External safety testing and red-teaming |
- inkling 2 small cost per task
- **About $1.18 per task on Artificial Analysis — far cheaper than closed models of similar score, but roughly six times its predecessor.** | Model | Cost per task | |---|---| | Inkling 2 Small | About $1.18 | | Gemini 3.8 Flash | About $7.77 | | Claude Sonnet 5.5 | About $10.02 |
The startup founded by OpenAI's former technology chief finally has a competitive model — though you can't download it yet. Thinking Machines Lab, founded by Mira Murati, introduced Inkling 2 Small on November 5, 2026, calling it "our most capable model, and the first of our second generation, with a big jump in agentic coding and vision."
The numbers back that up. On the Artificial Analysis Intelligence Index, Inkling 2 Small scores 31. Its predecessor, Inkling Small, scored 13.
1. The model — 594 billion parameters, 17 billion awake
| Item | Inkling 2 Small |
|---|---|
| Architecture | Sparse mixture-of-experts (MoE) |
| Total parameters | 594 billion |
| Active parameters | 17 billion per token |
| Inputs | Text and images (no audio) |
| Reasoning | Adjustable depth |
| Training | About 12 trillion tokens overall |
| Reinforcement learning | About 3 trillion tokens on 2,048 GPUs over nine days |
| Access | Tinker and Tinker Playground |
| Weights | Closed at launch; Apache 2.0 promised |
Mixture-of-experts models split a large network into specialist sub-networks and activate only a few for each token — here about 3% of the total — which keeps inference cheap. Thinking Machines lists its own weaknesses: inefficient communication at times, imperfect tool choice in large codebases, and less polish for chat. The Decoder quotes the company saying Inkling models "are designed as base models for fine-tuning" — raw material for its Tinker fine-tuning platform rather than a finished chatbot.
2. Where it stands — a Western open-model leader, if the weights ship
| Model | Developer (country) | AA index | Weights |
|---|---|---|---|
| Claude Opus 5.5 | Anthropic (US) | 58 | Closed |
| GPT-6 Astra | OpenAI (US) | 57 | Closed |
| Kimi K3.1 Pro | Moonshot AI (China) | 46 | Open |
| DeepSeek V4.1 Flash | DeepSeek (China) | 32 | Open |
| GLM-5.3 | Z.ai (China) | 32 | Open |
| Inkling 2 Small | Thinking Machines (US) | 31 | Promised |
| Mistral Large 4 | Mistral (France) | 26 | Open |
| Meta Muse Spark Open | Meta (US) | 25 | Open |
Every open model above it is Chinese. Among Western open models it beats October's trillion-parameter Mistral Large 4 and Meta's Muse Spark Open. It still trails closed frontier models by 27 points. Artificial Analysis says it will withhold the "strongest open model outside China" label until the weights are actually public.
The detailed results are uneven. Inkling 2 Small led its comparison set on visual document understanding (GDP.pdf, 25.2%) and perception (PerceptionBench, 64.7%), but trailed DeepSeek V4.1 Flash on agentic coding (FrontierCode 2.0: 25.4% vs 28.5%; DeepSWE: 72.1% vs 77%). It costs about $1.18 per Artificial Analysis task, versus about $7.77 for Gemini 3.8 Flash at a similar score.
3. What remains — weights, money and the content of 'safety testing'
| Open question | What is known |
|---|---|
| Weights release date | 'Coming weeks' per the company; no date |
| Safety testing | External testing and red-teaming; criteria not disclosed |
| Funding | Talks for $5–6 billion at $40 billion+ pre-money reported, unconfirmed |
| Prior funding | $2 billion seed at $12 billion valuation, July 2025 |
The Information reported that Thinking Machines is negotiating a $5–6 billion round led by Accel, with Nvidia possibly providing about half. An earlier attempt at a $50–60 billion valuation collapsed this year, and three co-founders returned to OpenAI. This model is the company's first report card since.
Once weights are released, a company can no longer control how its model is used. Whether Thinking Machines publishes the criteria and results of its safety testing alongside the weights will shape how much trust it earns — and independent developers will then verify the benchmarks for themselves.
Sources
- The Decoder — Thinking Machines' Inkling 2 Small claims top spot among western open-weight models (2026-11-06)
- Thinking Machines Lab — Inkling (official model page)
- Trending Topics — Inkling 2: Thinking Machines Overtakes Mistral in Open Weights (2026-11)
- MarkTechPost — Thinking Machines Releases Inkling 2 Small, a 594B MoE With 17B Active (2026-11-05)
- The Next Web — Thinking Machines is in talks to raise $6bn, The Information reports