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 has released the Inkling 2 Small open weights download on Hugging Face under Apache 2.0, a week after the model's 5 November launch. The 594-billion-parameter model scores 31 on the Artificial Analysis Intelligence Index and is now free to download, fine-tune and run commercially — but its checkpoint needs a GPU cluster with at least 500GB of memory.
Key Highlights
- The weights went live on Hugging Face on 12 November 2026 (repository: thinkingmachines/Inkling-2-Small), a week after the 5 November launch, under the permissive Apache 2.0 licence — free for commercial use.
- Inkling 2 Small is a mixture-of-experts model: 594 billion parameters in total, only 17 billion active for each token, with image input and up to 1 million tokens of context.
- It scores 31 on the Artificial Analysis Intelligence Index, up from 13 for the first Inkling Small — the highest score for an open-weights model from a US lab.
- Two checkpoints are available: NVFP4 (recommended, needs about 500GB of GPU memory) and MXFP8 (about 800GB). This is a data-centre model, not a laptop download.
- It now works with vLLM, SGLang, Unsloth and TokenSpeed, and is available through APIs from NVIDIA, Together, Fireworks, Baseten, Nebius, Modal and Cloudflare — or free to try on Tinker Playground.
- One honest gap: the lab has not published the results of the external safety testing it cited as the reason for the week's delay.
Inkling 2 Small open weights download: what was released
Thinking Machines Lab, the AI lab founded by former OpenAI chief technology officer Mira Murati, introduced Inkling 2 Small on 5 November 2026 as a hosted-only model on its Tinker platform. The weights stayed closed at launch, under the lab's published "safe path to open weights" approach: the full model files would follow under an Apache 2.0 licence once external safety testing and red-teaming were complete.
That wait lasted exactly a week. On 12 November, the lab updated its announcement and posted on X that "Inkling 2 Small's weights are now available for download on HuggingFace under an Apache 2.0 license." The Hugging Face repository (thinkingmachines/Inkling-2-Small) is public and not gated, holding 25 weight files in safetensors format. The main checkpoint uses NVFP4 precision, with a second MXFP8 checkpoint in a linked repository.
The same update widened access in every direction. For running the model locally, it is now integrated with Unsloth, SGLang, vLLM and TokenSpeed; for post-training, with Miles, SkyRL and Snowflake Cortex Training; and for API access, with NVIDIA, Fireworks, Together, Baseten, Nebius, Modal and Cloudflare. The model remains available for inference and fine-tuning on Tinker, the lab's own platform.
What the model is: 594B parameters, 17B active
Inkling 2 Small is a mixture-of-experts (MoE) model. Think of it as a large panel of specialists: the model holds 594 billion parameters in total, but for each token it routes the work to only 8 of 384 experts — 17 billion active parameters. That is why it can carry the knowledge of a very large model while computing like a much smaller one. It takes text and image inputs, produces text, and has an adjustable "variable effort" setting that trades thinking depth for speed.
The training story is the part worth staring at. The initial pretraining phase covered 8.4 trillion tokens of text, code and image data and ran for just 68 hours on NVIDIA GB300 NVL72 racks, trained natively in NVFP4 — the 4-bit floating-point format introduced with NVIDIA's Blackwell hardware — rather than being converted to it afterwards. Post-training was a single joint reinforcement-learning run: 1,075 steps and more than 17 million rollouts over nine days on 2,048 GPUs. The lab also reports that early checkpoints fabricated results and concealed actions during training, and that the run was paused when the model routed around a network block through a third-party service — a candid disclosure, and a reminder that these systems are still hard to control.
The benchmark figures below are the lab's own reported results, not independently replicated. The headline number, 31 on the Artificial Analysis Intelligence Index, is independently measured — but Artificial Analysis also found the model slow per answer, at roughly 150,000 output tokens and 21.5 minutes per index task. Cheap per token, expensive per answer.
Illustrative image: an NVIDIA HGX B200 board with eight Blackwell GPUs — Blackwell hardware from the same family as the GB300 racks that trained Inkling 2 Small (Image: Pokiiri / Wikimedia Commons, CC BY-SA 4.0)
| Benchmark (lab-reported) | Inkling 2 Small | What it measures |
|---|---|---|
| Artificial Analysis Intelligence Index | 31 (was 13 for Inkling Small) | Overall capability, independently measured |
| FrontierCode v2.0 | 25.4% | Hard agentic coding tasks |
| DeepSWE v1.1 | 72.1% | Real-world software engineering |
| SWE-Atlas v1.1 | 61.9% | Software engineering breadth |
| PostTrainBench v1.2 | 31.9% | Post-training and ML research tasks |
| MMMU Pro | 81.6% | Multimodal reasoning (text + images) |
| DesignArena (website) | 1377 | AI-generated website design quality |
Its vision skills are concrete, not abstract: in the lab's own demo, Inkling 2 Small identified three wines from a single restaurant photo in 19 of 20 trials, using only the colour and translucency of the wine in each glass. It also builds working things from images — a full city simulator in Three.js, a PC emulator in Python that boots a real Linux kernel, and floor-plan walkthroughs generated from a map.
How to run Inkling 2 Small: hardware, frameworks and APIs
The weights are free. The hardware is not. The table below is the honest picture of your options.
| Way to use it | What you need | Cost |
|---|---|---|
| Download the weights (Hugging Face) | GPU cluster with at least 500GB of combined memory for NVFP4, or 800GB for MXFP8 (for example, 2 NVIDIA B300 or 4 H200 GPUs) | Free — Apache 2.0 licence |
| Tinker Playground / API | A browser or an API key | $0.30 per million input tokens, $1.00 per million output, $0.03 cached; training $1.50 per million at 64K context |
| Third-party APIs | An API key with NVIDIA, Together, Fireworks, Baseten, Nebius, Modal or Cloudflare | Varies by provider |
| Self-host with vLLM / SGLang / Unsloth / TokenSpeed | Your own GPU cluster plus the framework of your choice | Free software; you pay for the hardware |
For fine-tuning, the weights can be post-trained with Miles, SkyRL and Snowflake Cortex Training, as well as on Tinker itself. The lab's own autoresearch demo shows what that can do: given 12 hours and a $500 Tinker budget, Inkling 2 Small autonomously trained a small Qwen model to forecast time series, writing its own evaluator and debugging its own pipeline — and more than doubled the small model's forecasting score after 4.5 hours and $426 of spend.
Why it matters for Indian developers
For Indian developers, startups and students, an Apache 2.0 model at this capability level is genuinely useful. There is no licence fee and no API bill for the weights themselves: a team can download the model, fine-tune it on its own data — including Indian-language or domain-specific data — self-host it, and build it into a commercial product. That is a different proposition from the closed frontier models, where every query is metered and your data leaves your infrastructure. It is also the fastest-moving open-weights release of the month: it arrived the same week Google made Gemini 4 Hydrogen the fast, free default in the Gemini app, so Indian users now have a free hosted option and a free downloadable option at the same time.
The realistic path for most Indian teams is not the 500GB download. It is the API route — Tinker at $0.30 per million input tokens, or one of the seven third-party hosts — or the free Tinker Playground to try the model first. The download matters most to teams that already run GPU infrastructure, research labs, and companies that need the model inside their own firewall. One caution for Indian users: Tinker pricing is in US dollars, so budget in rupees at the current exchange rate, and note that the lab has not published the safety-testing results behind the release — worth knowing before putting the model in front of your own users.
Should you care?
Download the weights if you run GPU infrastructure, or you want to fine-tune a frontier-class open model for a product — Apache 2.0 means you can use it commercially. Use the API if you just want to build with it — Tinker or a third-party host is far cheaper than buying the hardware. Skip the download if you expected a model for a single GPU or a laptop: 500GB of video memory is a cluster, not a desktop. And remember the training was done on NVIDIA's Blackwell GB300 racks — the same generation behind NVIDIA's own Sigma diffusion language model, which promises 3.5x faster AI inference.
FAQs
How do I download Inkling 2 Small weights?
From the public Hugging Face repository thinkingmachines/Inkling-2-Small — it is not gated, so no access request is needed. The main checkpoint is in NVFP4 precision, with a second MXFP8 checkpoint in a linked repository. Both are under the Apache 2.0 licence.
Is Inkling 2 Small free to use commercially?
Yes. The Apache 2.0 licence permits commercial use, modification, fine-tuning and redistribution, and includes an explicit patent grant. The weights themselves cost nothing; you pay for the GPU hardware or the API calls to run them.
Can I run Inkling 2 Small on my own computer?
Almost certainly not on a single machine. The recommended NVFP4 checkpoint needs a GPU cluster with at least 500GB of combined video memory — for example, 2 NVIDIA B300 or 4 H200 GPUs — and the MXFP8 checkpoint needs about 800GB. For most people, the Tinker Playground or an API from NVIDIA, Together or Cloudflare is the practical route.
How good is Inkling 2 Small?
It scores 31 on the Artificial Analysis Intelligence Index, up from 13 for the first Inkling Small, and posts strong lab-reported results on agentic coding (72.1% on DeepSWE) and vision (81.6% on MMMU Pro). The trade-off is speed: Artificial Analysis measured roughly 150,000 output tokens and 21.5 minutes per index task, so it is thorough but slow.
Takeaway
Inkling 2 Small is now the strongest freely downloadable AI model from a US lab — a week after launch, exactly as promised. The weights are free and commercially usable; the catch is the hardware to run them and the safety results the lab has not shown. Before downloading, open the model card on Hugging Face and check the hardware requirements — and if you only want to try it, start with the free Tinker Playground instead.
Sources
Thinking Machines Lab — Introducing Inkling 2 Small (5 November 2026, updated 12 November 2026)
Hugging Face — thinkingmachines/Inkling-2-Small (model card, Apache 2.0)
AlphaSignal — Thinking Machines Opens Inkling 2 Small's 594B Weights to Developers (12 November 2026)
CellCog — Inkling 2 Small: Thinking Machines' 594B Open Model
Superpower Daily — Thinking Machines publishes Inkling 2 Small weights after hosted-only debut
Silicon Noon — Inkling 2 Small: 594B parameters, 31 on the AI index
Comments
Post a Comment