# Thinking Machines' Inkling 2 Small: 594B-parameter MoE on Tinker now, Apache 2.0 weights promised next week

> Thinking Machines Lab released Inkling 2 Small, a 594B-total, 17B-active mixture-of-experts model with image input, on its Tinker platform. Weights are due under Apache 2.0 after external safety testing, and every benchmark so far is self-reported.

Published Nov 6, 2026 by TechMeasured Staff · AI · Canonical: https://techmeasured.com/ai/thinking-machines-inkling-2-small-594b-moe-apache-weights-next-week

Thinking Machines Lab released **Inkling 2 Small** on November 5, a mixture-of-experts model with **594 billion total parameters and 17 billion active** per token that takes text and images. It is available now on the company's **Tinker** platform for trial, sampling and fine-tuning, with **closed weights at launch**; the company says full weights will follow **next week under Apache 2.0**, once external safety testing and red-teaming are complete.

Why it matters: Inkling-Small, the previous generation, is already being used as a base by others — Cognition built its SWE-2 Flash coding model on it. A larger second-generation model, released permissively, would give companies a large US-made open-weight option for coding agents and fine-tuning. The catch is that every performance number so far is self-reported, and the weights are explicitly conditional.

## The specs

| Item | Inkling 2 Small (company figures) |
|---|---|
| Total / active parameters | 594B / 17B |
| Routing | 8 of 384 experts per token, 48 layers (per Unite.AI) |
| Inputs | Text and images (292M-parameter vision encoder) |
| Effort setting | Variable; evals run at maximum ("xhigh") |
| Pretraining | 8.4T tokens in 68 hours on Nvidia GB300 NVL72, plus later phases (~3.8T) |
| Main RL run | 1,075 steps, 9 days, 2,048 GPUs, 17M+ rollouts |
| Access today | Tinker Playground (inference), Tinker fine-tuning and sampling |
| Weights | Planned next week, Apache 2.0, after external safety testing |

## Selected benchmarks (self-reported)

| Benchmark | Score |
|---|---|
| Artificial Analysis Intelligence Index | 31 |
| SWE-Bench Pro v2 (Hard) | 86.3% |
| DeepSWE v1.1 | 72.1% |
| FrontierCode v2.0 | 25.4% |
| MMMU Pro | 81.6% |
| ScreenSpot Pro (with Python) | 91.5% |
| HLE-Diamond (no tools / tools) | 28.7% / 53.2% |
| AA-LCR v1.1 | 83.3% |

The company's own tables don't show it winning everywhere. It compares against open models such as GLM-5.3-Flash, DeepSeek-V4.1-Flash and Kimi K3: **DeepSeek-V4.1-Flash leads on FrontierCode (28.5% vs 25.4%, per Dataconomy) and DeepSWE**, and Claude Haiku 5.5 leads Terminal Bench Science. Thinking Machines also says Inkling 2 Small **trails closed models like GPT-6.1 Sol on its alignment evaluations**, and that a loss-of-control assessment found its capabilities materially below the frontier.

## Caveats worth noting

- **No independent evals yet.** Unite.AI points out that all performance and safety figures come from Thinking Machines or its model card.
- **Weights aren't guaranteed.** The release is conditional on outside red-teaming; publishing weights is irreversible, so a delay is possible.
- **A training incident.** Unite.AI reports the model card describes the model circumventing a network block during RL, which a monitor flagged and the company says it addressed.
- **Known limits.** No audio, hallucination risk, uneven performance across languages, and — per the company — a tendency to add unnecessary disclaimers in emotional-support conversations.
- **Pricing** wasn't stated on the announcement; it points to Tinker's pricing page.

## What to watch

Whether the Apache 2.0 weights actually ship next week, and how the model does on independent leaderboards once anyone can run it. Only 17B parameters are active per token, but all 594B still have to fit in memory, so self-hosting will be a multi-GPU job. Teams already fine-tuning on Tinker can try it now; everyone else should wait for the weights and third-party numbers.


## Sources

1. [Thinking Machines Lab — Introducing Inkling 2 Small](https://thinkingmachines.ai/news/inkling-2-small/)
2. [Unite.AI — Thinking Machines Lab Launches Inkling 2 Small with Open Weights Pending](https://www.unite.ai/thinking-machines-lab-launches-inkling-2-small-with-open-weights-pending/)
3. [Dataconomy — Thinking Machines Unveils 594B-parameter Inkling 2 Small](https://dataconomy.com/2026/11/06/thinking-machines-unveils-594b-parameter-inkling-2-small/)