Thinking Machines’ Inkling 2 Small scores 31 on Artificial Analysis, with Apache weights promised next week
The Thursday post describes a mixture-of-experts model with 594 billion parameters and 17 billion active, on Tinker with closed weights. Artificial Analysis puts the same score on index v5.0.1 and says an open-weight release would lead models trained outside China. It also says the release date is unconfirmed. The lab’s own model card prints 30.6.
Feature4 minIn-depth
James Okonkwo4 min read
- Thinking Machines
- Open weights
- Inkling
- Tinker

Photo: Thinking Machines LabSource page
Thinking Machines Lab on Thursday introduced Inkling 2 Small, a mixture-of-experts model with 594 billion parameters in total and 17 billion active, taking images as well as text and offering what the company calls variable effort. The weights are closed for now. The post, dated 5 November, says the model is available on Tinker for sampling and fine-tuning, and on Tinker Playground, and that checkpoints will be downloadable when the weights ship. The company says it intends to release the full weights under an Apache 2.0 license next week, after external safety testing and red-teaming. The model card lists the current license as the Tinker terms of use.
On the Artificial Analysis Intelligence Index, the company’s post says Inkling 2 Small scores 31, well above its previous models, with the gains in agentic coding, vision, knowledge work, and reasoning. Artificial Analysis, in its own write-up the same day, uses that 31 for index version 5.0.1 and says it is 18 points above Inkling Small. On current rankings, Artificial Analysis says, releasing the weights would make this the highest-scoring open-weight model developed outside China. It also lists the Apache release as planned with the date unconfirmed, which is a looser schedule than “next week.” The lab’s model card, in a comparison table headed AA Index (v5.0), prints 30.6 for Inkling 2 Small, against 32.6 for DeepSeek-V4.1-Flash-1017 and 36.5 for Claude Haiku 5.5. The 31 and the 30.6 are not the same printed figure. They come from the company’s narrative, from Artificial Analysis’s v5.0.1 page, and from the model card’s v5.0 column.
What the index says it costs
Artificial Analysis prices the model at $0.30 per million input tokens and $1.00 per million output tokens, with cached input at $0.03. It says a task on its Intelligence Index costs $1.18, against $0.20 for Inkling Small, $7.77 for Gemini 3.8 Flash at its high setting, and $10.02 for Claude Sonnet 5.5 at medium. The token bill is higher in part because the model writes more. Artificial Analysis puts average output at about 150,000 tokens per index task, about four times Inkling Small’s 37,000, and above Gemini 3.8 Flash high at about 99,000 and Sonnet 5.5 medium at about 31,000. On FrontierCode 2.0 it scores 25 percent, against 11 percent for Sonnet 5.5 medium and 5 percent for Inkling Small. On Terminal-Bench 4.0 it scores 13 percent, against 30 percent for that Sonnet setting. Hallucinations on AA-Omniscience v1.1 fall from 61 percent for Inkling Small to 27 percent, Artificial Analysis says, while the rate at which the model attempts an answer falls from 77 percent to 53 percent and accuracy rises only from 32 percent to 34 percent.
The training account in the lab’s post is specific about scale. A first pretraining phase covered 8.4 trillion tokens and ran for 68 hours on Nvidia GB300 NVL72 racks, natively in NVFP4, the 4-bit format Blackwell introduced. A second phase used 3.8 trillion tokens of newer data at a 512,000-token context. A third phase was about 320 billion tokens of reasoning and agent trajectories at a 1 million-token context, mixed with replayed pretraining data. A fourth kept about 34 billion tokens of the highest-quality agent trajectories. The company says those last two mixtures were mostly trajectories sampled from earlier models, including Kimi K3 and its own, filtered for correctness and alignment. For the reinforcement-learning run it says it uses a 256,000-token context and compaction for longer tasks. Evals, it says, allow a trajectory of up to 1 million tokens, up from 256,000 for Inkling Small, at effort 0.99 and temperature 1.0. Images go through a 292-million-parameter vision encoder, with 20-by-20-pixel patches and 2-by-2 pooling.
A hold on the weights, and the failures the lab describes
The safety section says the release of the weights is pending further evaluations by the lab and its partners. Early checkpoints, it says, were persistent on hard tasks and in some cases fabricated results, ignored instructions, and concealed what they had done. The company describes filters, an alignment reward during reinforcement learning, environments that reward giving up on impossible tasks, and self-distillation from a teacher prompted with a conduct spec. On an internal misbehavior score where lower is better, its table lists Inkling 2 Small at 14.9 and GPT-6.1 Sol at 0.8. It says the new model still trails closed models such as GPT-6.1 Sol, and that closing the gap is a priority. On chemical, biological, radiological, and nuclear risk, it says Inkling 2 Small’s knowledge and capability scores sit below the open-weight frontier. On Scale AI’s FORTRESS benchmark, the table’s average risk score for Inkling 2 Small is 16.2, which the text says is lower than Inkling and Inkling Small; Muse Spark Open, in the same table, is at 7.6.
One demonstration is an internal forecasting task. The post says Inkling 2 Small trained a Qwen3.5-4B model on Tinker to predict time series such as weather and Wikipedia page views, writing its own evaluator and trying supervised fine-tuning and reinforcement learning. After 4.5 hours and $426 of Tinker spend, the lab says, the trained model had more than doubled its forecasting score. The task prompt printed beside that account allows 12 hours and $500. The limitations section is blunt about ordinary chat. The model, the company writes, is less optimized for it: worse organization and presentation, too many slashes, and an overreaction to mental-health and other safety queries. A footnote on DistressBench says the same thing in the other direction: the model handles distress well and sometimes adds disclaimers to ordinary emotional-support questions.


