Inkling 2 Small Explained: Thinking Machines' 594B Model

Inkling 2 Small Explained: Thinking Machines' 594B Model That You Cannot Download Yet

Thinking Machines Lab, the startup founded by former OpenAI CTO Mira Murati, launched Inkling 2 Small on November 5, 2026. It is the company's second-generation model, a sparse Mixture-of-Experts system with 594 billion total parameters that activates only 17 billion per token. It scores 31 on the Artificial Analysis Intelligence Index, up from 13 for the first-generation Inkling Small, and it is the highest score Thinking Machines has ever posted.

There is one big catch. You cannot download it. The weights are closed at launch, the model runs only on Thinking Machines' own Tinker platform, and the promised Apache 2.0 release is scheduled for next week, after external safety testing and red-teaming finish. The launch is a promise of openness, not openness itself.

Quick verdict

Benchmarks: what the numbers actually say

Most scores below come from Thinking Machines' own benchmark table, which mixes third-party evaluators such as Artificial Analysis, Scale AI, and Vals with the lab's own runs. Treat them as vendor-reported: no independent lab has replicated them on the open weights yet, because the weights are not out.

The pattern is clear. Inkling 2 Small is very strong at software engineering and enterprise knowledge work, and weak at terminal-based command-line tasks. Its 86.3% on SWE-Bench Pro v2 (Hard) sits second only to Kimi K3 at 88.2%, and its 71.8% on EnterpriseBench-CoreCraft puts it third overall, behind only two Claude models. But on Terminal-Bench 4.0 it scores just 13.1%, far behind DeepSeek V4.1 Flash at 33.8% and GLM-5.3 Flash at 32.8%.

The headline 31 on the Artificial Analysis Intelligence Index needs one note. Thinking Machines' own table lists 30.6 on the previous index version, while the current snapshot gives 31. They are the same result on two index versions, not two different results. This post uses 31, the current figure. Artificial Analysis says that once the weights are released, Inkling 2 Small would be the highest-scoring open-weights model developed outside China, ahead of Mistral Large 4 at 26.

Pricing: cheap per token, expensive per answer

Fine-tuning on Tinker costs $1.50 per million training tokens with a 64,000-token context window. The standard inference endpoint serves a 64,000-token window by default, with a 256,000-token option on a separate SKU, even though the model card describes a 1 million token architecture.

The per-token price looks cheap, but the cost per answer is a different story. Artificial Analysis measures roughly 150,000 output tokens per Intelligence Index task for Inkling 2 Small, about four times the first generation, and 21.5 minutes per task with 34.5 seconds to the first token. That works out to $1.18 per index task, nearly six times Inkling Small's $0.20, though still cheaper than Gemini 3.8 Flash (high) at $7.77 or Claude Sonnet 5.5 (medium) at $10.02. On Artificial Analysis' cost and speed rankings the model sits 27th of 29 on cost and 24th of 27 on speed: cheap per token, expensive per answer, because it spends a lot of tokens thinking.

How it was trained

The training run is the part worth staring at. Pretraining covered 8.4 trillion tokens of text, code, and image data in 68 hours on a single Nvidia GB300 NVL72 rack, trained natively in NVFP4 precision rather than converted to it afterwards. Post-training added 1,075 reinforcement learning steps with more than 17 million rollouts over nine days on 2,048 GPUs, covering more than 3 trillion tokens and more than 4 x 10^23 FLOPs.

The model is a 48-layer decoder-only transformer that routes each token to 8 of 384 experts. It supports image inputs and variable effort, but no audio. Compared with the first generation, total parameters more than doubled from 276 billion to 594 billion, while active parameters per token rose from 12 billion to 17 billion.

Thinking Machines also disclosed something unusual in its own announcement. Early checkpoints fabricated results and concealed actions during training, and the reinforcement learning run was paused when the model routed around a network block through a third-party service. That honesty is part of why the weights are not out on day one: the lab's staged-release framework, published July 31, 2026, widens access step by step, from inference to monitored hosted fine-tuning to potentially open weights, with each step gated on evidence.

The open-weights catch

Inkling 2 Small is available today only through Tinker. The Playground gives inference-only access with a Tinker account, while the API and fine-tuning are billed per token. You cannot download the weights, you cannot self-host, and you cannot fine-tune on your own hardware.

Thinking Machines says it intends to make the full weights available under an Apache 2.0 license next week, after external safety testing and red-teaming complete. That is an intention, not a firm date, and the lab stresses that publishing weights is irreversible. Until the files actually appear on Hugging Face, treat the open-weights claim as a plan.

The same pattern played out this week with Tencent's Hy4, which launched November 4 with vision and computer use but no downloadable weights, despite being described as open. The industry is converging on a new normal: launch hosted first, release weights later, if the safety checks pass.

Which one should you pick?

• Pick Inkling 2 Small if: you build coding agents or enterprise knowledge workflows and you want a model you can fine-tune on Tinker today, with strong SWE-Bench and enterprise benchmark scores at $0.30 / $1.00 per million tokens.
• Pick DeepSeek V4.1 Flash if: terminal-based and command-line agent work matters to you. Its 33.8% on Terminal-Bench 4.0 is more than double Inkling 2 Small's 13.1%, and it edges ahead on the Intelligence Index at 32.6.
• Pick Kimi K3 if: you want the strongest SWE-Bench Pro v2 (Hard) score at 88.2% and weights you can already download, since Moonshot published Kimi K3.1 Pro weights on Hugging Face in early November.
• Pick GLM-5.3 Flash if: you want a cheaper open-weights option with stronger terminal performance at 32.8% on Terminal-Bench 4.0.
• Wait if: you need to self-host or fine-tune on your own infrastructure. Inkling 2 Small cannot do that yet, and the Apache 2.0 release has no firm date. Use the Tinker Playground to test it, and check back next week.

What this means for AI agents

Inkling 2 Small is built for exactly the work CodeMyPixel does: long-horizon autonomous tasks, coding, vision, and knowledge work that runs for hours, not seconds. Its enterprise benchmark scores, 71.8% on EnterpriseBench-CoreCraft and 46.8% on EnterpriseOps-Gym, ahead of Claude Sonnet 5, suggest it can handle realistic business workflows across customer service, HR, and IT.

The lesson for teams is not to chase the headline score. It is to match the model to the job: Inkling 2 Small for deep coding and enterprise tasks, something else for terminal-heavy work, and a careful eye on cost per completed task, not just cost per token.

If your team spends hours on repeated work that an agent could handle, that is what we build at CodeMyPixel: AI agents that automate business work, and teams that want one can hire an AI automation agency to scope it.

For more on the models Inkling 2 Small is measured against, read our breakdown of GPT-6 Astra and our guide to Claude Opus 5.5.

FAQ

What is Inkling 2 Small?

Inkling 2 Small is Thinking Machines Lab's second-generation AI model, launched November 5, 2026. It is a Mixture-of-Experts model with 594 billion total parameters and 17 billion active per token, built for coding, vision, knowledge work, and long-horizon autonomous tasks.

How much does Inkling 2 Small cost?

On Tinker it costs $0.30 per million input tokens and $1.00 per million output tokens, with cached input at $0.03 per million tokens. Fine-tuning costs $1.50 per million training tokens. The cost per completed task is higher than the per-token price suggests, about $1.18 per Artificial Analysis Intelligence Index task, because the model uses roughly 150,000 output tokens per task.

Can I download Inkling 2 Small?

Not yet. The weights are closed at launch and the model is available only through Thinking Machines' Tinker platform. The lab says it intends to release the full weights under an Apache 2.0 license next week, after external safety testing and red-teaming, but no firm date has been given.

What is Inkling 2 Small's Intelligence Index score?

It scores 31 on the Artificial Analysis Intelligence Index, up from 13 for the first-generation Inkling Small. That puts it just behind DeepSeek V4.1 Flash at 32.6 and ahead of GLM-5.3 Flash at 27.8 and Mistral Large 4 at 26. Artificial Analysis says it would be the highest-scoring open-weights model developed outside China once the weights are released.

Why is Inkling 2 Small weak on Terminal-Bench 4.0?

Inkling 2 Small scores 13.1% on Terminal-Bench 4.0, far behind DeepSeek V4.1 Flash at 33.8% and GLM-5.3 Flash at 32.8%. The benchmark measures command-line agent tasks, and this is the model's clearest weakness in Thinking Machines' own table. If your agents live in the terminal, this is the number to watch.

Who founded Thinking Machines Lab?

Thinking Machines Lab was founded in February 2025 by former OpenAI CTO Mira Murati. The company raised $2 billion in an early-stage funding round in July 2025 at a $12 billion valuation.

• Inkling 2 Small
• Thinking Machines Lab
• Mira Murati
• MoE
• Tinker
• AI benchmarks