Thinky Machines releases Inkling 2 Small, 594B MoE beating Muse Spark
Original title: [AINews] Inkling 2 Small: 17A-594B trained in 68 hours on 12T tokens, beating Muse Spark
Apache 2.0 weights ship next week after safety testing; ranks 31 on Artificial Analysis Index.
Thinky Machines released Inkling 2 Small, a 594B total parameters / 17B active multimodal MoE. Trained from scratch in under 3 days on 8.4T tokens in NVFP4, it supports 1M context and is available for fine-tuning on Tinker. Apache 2.0 weights are promised next week after external safety testing; it outperforms or matches prior models on multiple benchmarks including AutomationBench and HLE-Diamond.
Unconfirmed: Exact token count and training details were reported by team members with some figures corrected post-publication; weights date and safety testing status remain pending external verification.
- Pretrained in 68 hours on GB300s with native NVFP4 precision
- 1M context window and 1M max output tokens
- Apache 2.0 open weights coming next week after safety testing
- Scores 31 on Artificial Analysis Intelligence Index
- Customizability is a design target for fine-tuning
- Vendors
- Thinky MachinesTinker
- Organizations
- Artificial Analysis
- AI models
- Inkling 2 Small
From the original article
14 min read at latent.space
November 5, 2026 — Inside Inkling 2 Small’s training run and METR’s study of AI research at Anthropic.
By swyx .ai · latent.space
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.