Just in
AIGenie 4 Now Available to Google AI Ultra Subscribers in Project GenieAICloudflare Clairic Goes Open Weight, Adds 9B Clairic-flash ModelTechGTA 6 Is Out Now on PS5 and Xbox Series X|S, Priced From $79.99TechAlibaba Earnings Show 53% Cloud Growth in September QuarterAIGemini 4 Neon Benchmarks, Price and Who Can Use It NowTechPalo Alto Networks Earnings: Q1 Revenue Up 39% to $3.44 Billion
AI

Inkling 2 Small Weights Released on Hugging Face Under Apache 2.0

Thinking Machines Lab has posted the full Inkling 2 Small weights on Hugging Face under an Apache 2.0 license, a week after launching the model on Tinker with closed weights.

USA Publishing Group3 min read
Share
Close-up of a computer motherboard with screws, the kind of server hardware needed to run Inkling 2 Small
Photo: Stan Hutter / Unsplash

Key takeaways

  • Thinking Machines Lab put the Inkling 2 Small weights on Hugging Face on Nov. 12, one week after a Nov. 5 launch that offered the model only through Tinker.
  • The model card lists 594 billion total parameters with 17 billion active, a context window of up to 1 million tokens, and an Apache 2.0 license.
  • The recommended NVFP4 checkpoint needs at least 500GB of combined GPU memory, such as two Nvidia B300 GPUs, while the MXFP8 version needs 800GB.
In this story 4 sections
  1. From Tinker only to open weights in seven days
  2. What Inkling 2 Small is
  3. Hardware you need to run it
  4. How the company says it performs

Anyone can now download Inkling 2 Small, the second generation model from Thinking Machines Lab, after the company posted its full weights on Hugging Face on Nov. 12. The release came one week after the Nov. 5 launch, when the model was offered only through the company's Tinker platform with closed weights.

Both the Inkling 2 Small model card and the Hugging Face repository list the license as Apache 2.0, which allows commercial use, modification and redistribution. At launch, Thinking Machines said the weights would follow "after the completion of external safety testing and red-teaming." An update dated Nov. 12 at the top of the company's announcement says the weights are now available. The company did not give a separate date for when that outside testing ended.

From Tinker only to open weights in seven days

  1. Nov. 5: Thinking Machines introduces Inkling 2 Small on Tinker, its training and inference service, and says weights will come the following week.
  2. Nov. 12: The company adds an update saying the weights are on Hugging Face, in a main NVFP4 checkpoint and a second MXFP8 version.
  3. Now: The updated announcement names several cloud and inference partners that serve it through their own APIs, and lists open source tools such as vLLM and SGLang that support it.

What Inkling 2 Small is

The model card describes a 48 layer Mixture of Experts design that routes each token to 8 of its 384 experts. It counts 594 billion total parameters, with 17 billion active for any given token. By our math, that means about 2.9% of the model (17 divided by 594) does the work on each step, which is how large sparse models keep serving costs down.

Other details from the card:

  • Context window of up to 1 million tokens
  • Text and image input, with text output only; the earlier Inkling Small model also took audio, but the new card lists no audio input
  • Native NVFP4 precision; the company says every stage of training used that 4 bit format
  • Release date on the card: Nov. 5, 2026

The Hugging Face page shows a different headline size of 383 billion parameters for the main checkpoint. The repository's own model table still lists 594 billion total and 17 billion active, so the smaller number appears to reflect how Hugging Face measures the stored files rather than a different model. Thinking Machines has not explained the gap.

Hardware you need to run it

This is not a laptop model. Thinking Machines says the NVFP4 checkpoint calls for 500GB or more of pooled GPU memory across a cluster, for example two Nvidia (NASDAQ: NVDA) B300 GPUs or four H200 GPUs. The MXFP8 version calls for 800GB or more, such as four B300s or eight H200s.

We calculated that the NVFP4 build needs 37.5% less GPU memory than the MXFP8 build (300GB less, out of 800GB). The company recommends NVFP4 for every use case and says the MXFP8 build scores only one to two percentage points higher on its evaluations.

How the company says it performs

Thinking Machines reports a score of 30.6 on version 5.0 of the Artificial Analysis Intelligence Index. Among the agentic coding results it lists, the model scored 72.1% on DeepSWE v1.1 and 61.9% on SWE-Atlas v1.1. On reasoning, the company reports 96.5% on ARC-AGI-1 and 81.7% on ARC-AGI-2. These are the company's own figures and have not been checked by us.

The company also says the new model "is more capable across the board than our first generation." Its earlier Inkling Small card used an older version of the Artificial Analysis index, so the two scores cannot be compared directly.

Inkling 2 Small joins a busy week of model releases, including the Gemini 4 Hydrogen rollout from Alphabet (NASDAQ: GOOGL) unit Google and Anthropic's Claude Fable 5.2 launch. For more coverage, see our AI news section.