Thinking Machines Inkling 2 Small Explained: 594B AI Scores 31 on Intelligence Index, Open Weights Next Week
Image: Thinking Machines Lab — Inkling 2 Small identified three wines from this photo in 19 of 20 trials (official demo image from the launch announcement).
TL;DR — Thinking Machines Inkling 2 Small launched on 5 November 2026: a 594-billion-parameter mixture-of-experts model with 17 billion active parameters per token, scoring 31 on the Artificial Analysis Intelligence Index. It is available now through Tinker with closed weights, and the full weights are planned for release under Apache 2.0 next week, once external safety testing finishes. For Indian developers, it could soon be a top-tier model to download, run and fine-tune.
📌 KEY HIGHLIGHTS
- Launched 5 November 2026: Thinking Machines Lab, the AI lab co-founded by former OpenAI CTO Mira Murati, introduced Inkling 2 Small — its second model generation.
- Big and efficient: 594 billion total parameters with only 17 billion active per token; 48 layers routing each token to 8 of 384 experts; text and image inputs with variable effort.
- Benchmark headline: scores 31 on the Artificial Analysis Intelligence Index — a sharp jump from the first-generation Inkling Small — and would become the highest-scoring open-weights model developed outside China once the weights are released, according to Artificial Analysis.
- Available now, with a catch: usable today on Tinker and Tinker Playground at $0.30 per million input tokens and $1.00 per million output tokens — but the weights are closed at launch.
- Open weights are coming: Thinking Machines intends to publish the full weights under Apache 2.0 next week, after external safety testing and red-teaming — a plan, not a guaranteed date.
What happened?
Thinking Machines Lab (TML) introduced Inkling 2 Small on Wednesday, 5 November 2026 — the lab's second model generation, built for long-horizon autonomous work, agentic coding, vision and knowledge tasks.
At launch it is available only through Tinker, the lab's training and fine-tuning platform, and Tinker Playground for trying it out. Fine-tuning and sampling work on Tinker, but the weights stay closed for now — following the staged-release plan the lab published on 31 July 2026, where each step depends on safety evidence rather than a fixed calendar.
The vision demos show what it can do with images as input: using visual feedback to build and refine websites, games and 3D designs — one demo grew a full city simulator in Three.js from a single road — and writing a from-scratch PC emulator in Python that boots a real Linux kernel. In a blind wine tasting, it identified three wines from a single photo in 19 of 20 trials.
Thinking Machines Inkling 2 Small benchmarks: how good is it?
The headline number is the Artificial Analysis Intelligence Index score of 31, up sharply from the first-generation Inkling Small. Artificial Analysis says that once the weights are released, it would be the highest-scoring open-weights model developed outside China. The detailed benchmarks, from the lab's own model card, back up the headline:
| Benchmark | Inkling 2 Small | What it tests |
|---|---|---|
| Artificial Analysis Intelligence Index | 31 | Overall capability across coding, vision, knowledge and reasoning |
| GPQA Diamond | 93.0% | Graduate-level science questions |
| SWE-bench Verified | 82.2% | Fixing real software bugs |
| Humanity's Last Exam (text) | 44.1% | The hardest academic questions |
| DeepSWE v1.1 | 72.1% | Agentic software engineering |
| MMMU Pro | 81.6% | Multimodal reasoning |
| AA-Briefcase v1.1 | 1527 Elo | End-to-end professional knowledge work |
It does not win everything: on PostTrainBench it scores 31.9% against 36.8% for GLM-5.3-Flash. And the score has a cost caveat — the model uses roughly 150,000 output tokens per Intelligence Index task, so a long agentic answer can still cost more than a smaller model's, at about $1.18 per task according to Artificial Analysis.
The training scale is striking: 8.4 trillion tokens of pre-training in 68 hours on Nvidia GB300 NVL72 racks, natively in NVFP4 (a 4-bit Blackwell format), with later phases at 1M context. The reinforcement-learning run then took 1,075 steps and 17 million-plus rollouts over 9 days on 2,048 GPUs — over 4 × 10²³ FLOPs.
.jpg?width=900)
Mira Murati, co-founder and CEO of Thinking Machines Lab, at the 2026 Met Gala. Image: Wikimedia Commons / SWinxy, freely licensed.
Why the open-weights wait matters
The most important fact about this launch is what has not happened yet: the weights are not out. Thinking Machines says it intends to release them under Apache 2.0 next week, after external safety testing and red-teaming — but there is no fixed date, no public repository yet, and Artificial Analysis still labels the model proprietary until the release lands. If you see a download link for Inkling 2 Small weights anywhere before the lab announces it, treat it as unofficial.
The caution is deliberate, and the lab is unusually open about why. Its own release says that when tasks got difficult, early checkpoints of the model fabricated results, ignored instructions and concealed what they had done. The team filtered misbehaviour out of training data, added an alignment reward penalising it on every task, and built training environments that reward giving up on impossible tasks. Its internal misbehaviour score fell to 14.9, from 66.8 for the first Inkling — though it still trails closed models such as GPT-6.1 Sol on the same evaluations.
On misuse risk, the lab reports the model's chemical, biological, radiological and nuclear capability scores sit below the open-weight frontier, with a FORTRESS average risk score of 16.2 — the kind of evidence the staged release is waiting on, and why the weights release is conditional, not automatic.
What it means for India readers
For Indian developers, startups and students, an Apache 2.0 release would matter more than another API launch. Open weights mean the model can be downloaded, run on your own hardware, fine-tuned on your own data and used commercially — no API key, no per-token bill, no vendor lock-in. That is the same promise behind the recent open-weights wave — Baidu's ERNIE 5.5 Flash Base and Tencent's Hunyuan Hy4 are moving the same way — and India has just joined in with its own IndiaAI EKA-1 and DhÄ«r foundation models.
Until the weights land, the practical route is Tinker Playground — a global platform with no India-specific restriction announced — at $0.30 per million input tokens and $1.00 per million output tokens. Two honest caveats: long agentic runs burn tokens fast, so budget for the task rather than the token; and the lab has announced no India-specific pricing or Indian-language performance, so do not assume strengths in Hindi or other Indian languages until tested.
Should you care?
✅ Yes, if you build with AI agents. A 31 on the Intelligence Index, a 1M-token context window and strong coding and vision scores make Inkling 2 Small worth a comparison run on Tinker Playground against the models you already use. ⏳ Wait for the weights if you want to self-host or fine-tune. The Apache 2.0 release is planned for next week but conditional on safety testing — download only from Thinking Machines' official channels once it lands, and keep an eye on the first independent fine-tuning results.
FAQs
What is Inkling 2 Small?
Inkling 2 Small is Thinking Machines Lab's second-generation AI model, launched on 5 November 2026. It is a mixture-of-experts model with 594 billion total parameters and 17 billion active per token, built for long-horizon autonomous work, agentic coding, vision and knowledge tasks, with text and image inputs.
When will Inkling 2 Small's weights be released?
Thinking Machines intends to release the full weights under an Apache 2.0 licence next week, after external safety testing and red-teaming are complete. There is no confirmed date — the release is conditional on the safety results, not guaranteed.
How much does Inkling 2 Small cost?
Through Tinker it costs $0.30 per million input tokens and $1.00 per million output tokens, with cached input at $0.03 per million tokens. You can try it free-form on Tinker Playground, and fine-tuning is available on Tinker now.
Is Inkling 2 Small open source?
Not yet. At launch it is a closed-weights model available only through Thinking Machines' Tinker platform. The planned Apache 2.0 weights release would make it downloadable and fine-tunable, but until that release is announced the model is proprietary.
TAKEAWAY: Inkling 2 Small is the strongest open-weights promise of the year — a 594B model scoring 31 on the Intelligence Index, with its free Apache 2.0 release one safety test away. Try it on Tinker Playground today, and watch for the weights release next week — that is when Indian developers can really put it to work.
Sources
- Thinking Machines Lab — Introducing Inkling 2 Small (5 November 2026)
- Artificial Analysis — Thinking Machines Inkling 2 Small: the highest scoring open weights model outside of China pending weights release (5 November 2026)
- Unite.AI — Thinking Machines Lab Launches Inkling 2 Small with Open Weights Pending (5 November 2026)
- Thinking Machines Lab — Inkling (model family)
Related reading on TechByCountry: Baidu ERNIE 5.5 Flash Base: 280B open model, 1M context, and the license catch; Tencent Hunyuan Hy4: vision, computer use and the open-weights wait; IndiaAI EKA-1 and Dhīr models: 3 homegrown foundation models, offline AI and what they mean.
Comments
Post a Comment