Thinking Machines Releases Inkling 2 Small Weights a Week After Launch

Thinking Machines Lab has made the weights of Inkling 2 Small downloadable, a week after launching the 594-billion-parameter model with closed weights. The Hugging Face repository is not gated and carries an Apache 2.0 licence tag.

The lab announced Inkling 2 Small on Nov. 5 and said the weights would follow after further safety evaluations by it and its partners. An update dated Nov. 12 on the announcement page confirms the download. The Hugging Face listing shows 3,647 downloads and 304 likes as of Nov. 14.

What the Inkling 2 Small model card specifies

The card describes a 48-layer mixture-of-experts transformer with 594 billion total parameters and 17 billion active per token, routing each token to 8 of 384 experts. It takes text and images and returns text. The context window reaches 1 million tokens.

The model was trained natively in NVFP4, a 4-bit format. The card lists two checkpoints and their memory floors.

Checkpoint	Minimum aggregate VRAM	Example hardware
NVFP4	500 GB	2x B300 or 4x H200
MXFP8	800 GB	4x B300 or 8x H200

The card says the MXFP8 checkpoint gains only 1 to 2 percentage points on evaluations and recommends NVFP4.

How it compares to the July model

Thinking Machines released Inkling Small on July 30, a 276-billion-parameter model with 12 billion active, according to The Decoder. That model had a 256,000-token context window and an Apache 2.0 licence, The Decoder reported.

	Inkling Small (July)	Inkling 2 Small
Total parameters	276B	594B
Active parameters	12B	17B
Context window	256K	1M

The Decoder reported Artificial Analysis scored Inkling Small at 40 on its Intelligence Index. The new card cites 30.6 on AA Index v5.0, and neither page says whether the two index versions are comparable, so the drop should not be read as a regression.

Benchmarks and who ran them

The announcement says most third-party scores come from the benchmark owners' own reports, among them Artificial Analysis, Vals, Scale AI, ARC Prize and MathArena. Internal evaluations were run by Thinking Machines.

Figures from the card include 86.3% on SWEBench Pro v2 (Hard), 90.2% on BrowseComp with compaction, 81.7% on ARC-AGI-2 and 28.7% on HLE-Diamond. The same card reports 5.7% on Terminal Bench Science v0.1.

One claim that does not match its own table

The announcement describes an automated-research case study that used 4 hours 17 minutes and $426 of Tinker spend. It says a forecasting score "more than doubled." The accompanying table lists a skill score moving from 0.191 to 0.289, a rise of roughly 50%.

What the lab says it does not do well

The post concedes that Inkling 2 Small is less polished for chat, sometimes overreacts to safety-related queries and sometimes picks poor tools in codebases with many skills. Its alignment results trail closed-weight models such as GPT-6.1 Sol, according to the lab.

Training used 8.4 trillion tokens in pretraining and a joint reinforcement-learning run on 2,048 GPUs over nine days. The card links an Acceptable Use Policy and a training data document, and the Apache tag alone does not describe that policy's terms.

For comparison with other open releases, see Alibaba's Qwen3.8-27B card and Nous Research's Arete-27B. The lab has not published pricing for its hosted API in the announcement.