Thinking Machines Lab

Inkling 2 Small

Provider

Thinking Machines Lab

Status

available

Context

1,000,000 tok

SWE-bench

82.2%

Price

$0.3 / $1 /MTok

Inkling 2 Small is the second-generation model from Thinking Machines Lab, released on 5 November 2026: a mixture-of-experts model with 594 billion parameters, 17 billion active, that takes text and images and has a variable reasoning effort (Thinking Machines). Unlike the lab’s first models, Inkling and Inkling-Small, it launched with closed weights, on Thinking Machines’ Tinker service only. Thinking Machines says it intends “to make the full weights available under an Apache 2.0 license next week after the completion of external safety testing and red-teaming”.

Artificial Analysis scored it 31 on its Intelligence Index v5.0.1 at launch, level with Gemini 3.8 Flash, against 13 for Inkling-Small and 12 for Inkling on the same index. On Tinker it costs $0.30 per million input tokens and $1.00 per million output tokens (Tinker).

Quick specs

VIEW INKLING 2 SMALL →

What Inkling 2 Small is

Thinking Machines calls Inkling 2 Small “our second model generation”. Its model card describes a 48-layer decoder-only transformer with a sparse mixture-of-experts backbone, in which each token is routed to 8 of 384 experts, and attention that mixes local and global layers. Images pass through a 292-million-parameter vision encoder based on Perception Encoder, with image patches interleaved with text tokens. The model card lists text and image input and text output; Inkling and Inkling-Small also took audio.

Thinking Machines says the first pretraining phase covered 8.4 trillion tokens in 68 hours on NVIDIA GB300 NVL72 racks, trained natively in the NVFP4 format, and that its main reinforcement-learning stage was a single joint run of 1,075 steps and more than 17 million rollouts over nine days on 2,048 GPUs. For the first Inkling release it ran separate reasoning and agentic runs and combined them by distillation. The effort setting comes from training: the lab “randomly sampled and conditioned on an effort level, and applied a length penalty in the reward function”.

Benchmark performance

The independent view first.

Source: Artificial Analysis and its leaderboard, independent. Artificial Analysis describes the model as “notably slow per task and very verbose”. Thinking Machines’ launch model card gives its score as 30.6 on v5.0.

Thinking Machines’ own figures, from its launch post and model card, are a ceiling:

SWE-bench Pro v2 is a different version from the SWE-bench Pro scores quoted on other model pages, so the two are not comparable. On DesignArena’s website leaderboard, as Thinking Machines reproduces it for 5 November, Inkling 2 Small rates 1,340, behind Muse Wonder (1,366) and Muse Spark 1.3 (1,351). Thinking Machines also says the model reaches Inkling-Small’s final score in one of its reinforcement-learning fine-tuning tests after 75 training steps.

Price and access

Inkling 2 Small runs on Tinker, Thinking Machines’ service for sampling and fine-tuning, and in the Tinker Playground. Tinker charges $0.30 per million input tokens ($0.03 cached) and $1.00 per million output tokens, against $1.87/$4.68 for Inkling and $0.58/$1.44 for Inkling-Small on the same list. Fine-tuning costs $1.50 per million training tokens on the 64K-context version and $3.00 on the 256K version.

Thinking Machines says the model “will initially be available through Tinker”. Artificial Analysis lists one API provider, Thinking Machines itself, at the same token prices, and measures $1.18 per completed index task.

Weights and licence

At launch the weights are closed, and the model card gives the licence as the Tinker Terms of Use. Thinking Machines says it is following the approach set out in its post A safe path to open weights, and that it intends “to make the full weights available under an Apache 2.0 license next week after the completion of external safety testing and red-teaming”. Checkpoints fine-tuned on Tinker “will be downloadable with the release of Inkling 2 Small’s weights”. Artificial Analysis lists the model as proprietary.

Inkling and Inkling-Small were released with Apache 2.0 weights on Hugging Face.

How it compares

vs Inkling and Inkling-Small

At launch, Inkling 2 Small scored 31 on the Intelligence Index v5.0.1, against 13 for the 276-billion-parameter Inkling-Small and 12 for the 975-billion-parameter Inkling (Artificial Analysis). Thinking Machines says it “is more capable across the board than our first generation, particularly in long-horizon autonomous work”. It is cheaper per token on Tinker than either, but it takes no audio and its weights are not yet released. The Inkling page covers the first generation.

vs Gemini 3.8 Flash

When Inkling 2 Small launched, Gemini 3.8 Flash also scored 31 on Artificial Analysis’s index, at high reasoning. It costs $0.75/$3.75 per million tokens on Google’s Gemini API until 31 December, more per token than Inkling 2 Small’s $0.30/$1.00, and Google published no SWE-bench figure for it.

vs the open-weight models

At Inkling 2 Small’s launch, GLM-5.3 scored 32 on the same index, and it is open-weight under a custom licence; Mistral Large 4 scored 26 and Muse Spark Open 25, then the highest-scoring US-developed open-weight model. Inkling 2 Small’s 31 is above both, but until its weights are released it is not an open-weight model.

Known limitations

Closed weights and one host at launch. It runs only on Tinker until Thinking Machines releases the weights.

Slow and verbose. Artificial Analysis measured 21.2 minutes and 154 million output tokens to run its index, against a median of 89.8 million output tokens among the models it compares it with.

No audio. Its predecessors took text, images and audio; the model card lists text and images only.

A low Terminal-Bench 4.0 score. On Thinking Machines’ own model card it posts 13.1% on Terminal-Bench 4.0, against 41.9% for GLM-5.3.

The vendor’s own caveats. The model card lists hallucination, imprecise instruction-following and “degraded performance in long multi-turn conversations”. Thinking Machines says the model “sometimes overreacts with disclaimers in normal emotional support queries”, and that it “still trails closed-weights models, such as GPT-6.1 Sol on our alignment evaluations”.

Frequently asked questions

When was Inkling 2 Small released?

Thinking Machines released Inkling 2 Small on 5 November 2026, on its Tinker service and in the Tinker Playground, with closed weights.

Are Inkling 2 Small’s weights available?

Not at launch. The model card gives the licence as the Tinker Terms of Use, and Thinking Machines says it intends to release the full weights under Apache 2.0 after external safety testing and red-teaming. Its first models, Inkling and Inkling-Small, were released with Apache 2.0 weights.

How much does Inkling 2 Small cost?

On Tinker’s price list, $0.30 per million input tokens ($0.03 cached) and $1.00 per million output tokens, with fine-tuning at $1.50 per million training tokens on the 64K-context version. Artificial Analysis puts a completed Intelligence Index task at $1.18.

How many parameters does Inkling 2 Small have?

594 billion in total, with 17 billion active per token, according to Thinking Machines. Its model card says each token is routed to 8 of 384 experts.

How does Inkling 2 Small score on independent tests?

At launch, Artificial Analysis scored it 31 on its Intelligence Index v5.0.1, level with Gemini 3.8 Flash and a point under GLM-5.3, against 13 for Inkling-Small and 12 for Inkling. Artificial Analysis also describes it as slow per task and verbose, at 154 million output tokens to run the index.

Does Inkling 2 Small accept audio?

No. Thinking Machines’ model card lists text and image input and text output. Inkling and Inkling-Small, the lab’s first models, also accept audio.