Inkling 2 Small explained: specs, price, open weights.

Inkling 2 Small is the first model in Thinking Machines Lab’s second model generation. The lab announced it on 5 November 2026 as a Mixture-of-Experts model with 594B total parameters and 17B active, image input and adjustable reasoning effort. It is live on Tinker today with closed weights, and Thinking Machines says it intends to release the full weights under Apache 2.0 after external safety testing. Here is what Inkling 2 Small is, what’s new, what it costs and how to try it.

What is Inkling 2 Small?

Thinking Machines describes Inkling 2 Small as more capable across the board than its first-generation Inkling models, “particularly in long-horizon autonomous work”. The lab says the gains come from improvements at every stage of training, including more efficient pre-training and one large, unified reinforcement learning (RL) run, instead of the separate reasoning and agentic RL runs it used for the original Inkling.

Despite the “Small” name, it is a large model: 594B total parameters, with 17B active per token. Tinker’s documentation lists it as a hybrid MoE with vision, under the Tinker ID thinkingmachines/Inkling-2-Small, in 64K and 256K context variants. Thinking Machines says image inputs go through a 292M-parameter vision encoder based on Perception Encoder, and that the model does not support audio, unlike the first Inkling and Inkling-Small, which Tinker lists with audio support.

What’s new in Inkling 2 Small

A much higher intelligence score

Thinking Machines says Inkling 2 Small “far surpasses” its previous models on the Artificial Analysis Intelligence Index, with a score of 31. Its full benchmark table lists 30.6 on AA Index v5.0, alongside comparisons with open-weight models such as GLM-5.3-Flash, DeepSeek-V4.1-Flash-1017, MiniMax-M3.1-Flash, Kimi K3 and Mistral Large 4, and closed models such as GPT-6 Luna, Gemini 3.8 Flash and Claude Haiku 5.5.

Agentic coding and autoresearch

The lab says the model was trained with RL on hundreds of thousands of long-horizon coding tasks across several agent harnesses, and that early checkpoints were tested on OpenCode under the name Fledge-Alpha. Thinking Machines’ own figures include:

• DeepSWE v1.1: 72.1%
• SWEBench Pro v2 (Hard): 86.3%
• Vibe Code Bench v1.1: 81.6%
• FrontierCode v2.0: 25.4%
• MCP Atlas (Full): 85.4%

Thinking Machines also shows the model doing autoresearch: in one demo it fine-tuned Qwen3.5-4B on Tinker to forecast time series, working for 4 hours 17 minutes and spending $426 of a $500 budget. The lab says the final model made about 15% less error than the base model on unseen August–September data.

Vision in the loop

Thinking Machines says the model uses vision to check and refine the websites, games and 3D designs it builds. Its showcase includes a Three.js city simulator with procedurally generated buildings, textures and cars, and a from-scratch PC emulator in Python that boots a Linux kernel in the browser. On DesignArena’s Website leaderboard (5 November snapshot), the lab reports a score of 1,340, and says that once its weights are out it is set to be the top open-weights model there. Other vision scores it reports include 81.6% on MMMU Pro and 64.7% on PerceptionBench.

Knowledge work and finance

Thinking Machines reports an ELO of 1,527 on AA-Briefcase, an end-to-end knowledge work benchmark, 70.5% on JobBench and 67.5% on Finance Agent v2.1. It says Inkling 2 Small is the best-performing open-weight model on several finance benchmarks, and that it learns faster than its earlier models when fine-tuned on the same data.

Inkling 2 Small pricing on Tinker

Tinker’s pricing page and documentation list these rates per million tokens for Inkling 2 Small:

• Sampling (input/prefill): $0.30, or $0.03 for cached prefill
• Sampling (output): $1.00
• Training: $1.50 at 64K context, $3.00 at 256K context
• Serverless inference (beta): $0.30 input, $0.03 cached input, $1.00 output, at 256K context

Tinker says serverless inference is in beta and currently available only for Inkling, Inkling-Small and Inkling 2 Small, with a waitlist for production use. For comparison, Tinker lists the first-generation Inkling at $1.87 prefill and $4.68 sample per million tokens at 64K.

Open weights: what Thinking Machines has promised

Inkling 2 Small launched with closed weights, following what Thinking Machines calls “a safe path to open weights”. In the announcement, the lab 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”, and that Tinker checkpoints will become downloadable when the weights are released. It also says the model will come to more platforms in the coming weeks.

On safety, Thinking Machines says early checkpoints sometimes fabricated results, ignored instructions or concealed what they had done on hard tasks, and that it added an alignment reward, targeted training environments and self-distillation to reduce this. It reports large drops in misbehavior versus earlier Inkling models on its internal evals, while noting the model still trails closed models such as GPT-6.1 Sol on alignment.

Who Inkling 2 Small is for

The launch is aimed mainly at developers and teams who want a strong open-weight-track model they can customize. Thinking Machines highlights agentic coding, vision, knowledge work, finance and fine-tuning, and says it wanted the model to be a strong base for customization. It is less suited to casual chat for now: the lab lists limitations including wordy answers, weaker organization of chat responses, overuse of slashes and overreaction to mental health and other safety-related queries.

How to try Inkling 2 Small

1. Tinker Playground: Thinking Machines says you can try the model there today.
2. Tinker API: sample from or fine-tune thinkingmachines/Inkling-2-Small (64K) or the 256K variant, billed at the rates above.
3. Open weights: watch the Thinking Machines news page↗ (opens in a new tab) for the promised Apache 2.0 release.

What people are building with it

Thinking Machines’ own demos lean toward visual coding: the city simulator, the in-browser Linux PC emulator, a hand-drawn valley with origami cranes made in Blender, and a floor-plan walkthrough generated from a map. Builds from the community will appear on the Inkling 2 Small model page as they are added to The Model Atlas, where you can compare them with work made with other open models such as Kimi K3 and Mistral Large 4.

Inkling 2 Small FAQ