Inkling 2 Mini
Thinking Machines released Inkling 2 Mini, an open-weights looped mixture-of-experts model with 26.4B total parameters that runs on a laptop and scores 48.2% on DeepSWE v1.1.
Inkling 2 Mini runs its middle layers three times for every token, so it reasons more with fewer output tokens. Most weights were trained and deployed natively in NVFP4, with the last four layers in MXFP8, giving a memory footprint of about 20GB. The lab also released a research model that can loop one to five times so outsiders can study monitorability.
- Date
- Friday 20 November 2026
- Lab
- Thinking Machines
- Kind
- open-weights
- Access
- open weights
- Price
- $0.075 in / $0.30 out per M tokens on OpenRouter
Figures
| Measure | Value | Measured by |
|---|---|---|
| DeepSWE v1.1 tasks resolved | 48.2% lab says it leads open-weights models under 100B total parameters; Qwen3.8-27B scored 42.2% | company |
| Vals Index v2.2 | 46.5% | company |
| Artificial Analysis Intelligence Index | 20 matches Qwen3.8 27B; $0.59 and 9 minutes per task versus $3.53 and 38.2 minutes for Qwen | Artificial Analysis |
| Parameters | 26.4B total, 3.4B active MoE; 5.4B counting loops, per The Next Web | company |
The Next Web reports an Apache 2.0 licence and a 1M token context window. OpenRouter also lists a 1M token context.
Sources
- thinkingmachines.ai/news/inkling-2-mini/
- thenextweb.com/news/inkling-2-mini-thinking-machines-looped-open-weights-laptop
- openrouter.ai/thinkingmachines/inkling-2-mini
- github.com/ShinMegamiBoson/lookatmyquant
This record was partly confirmed: some claims could not be checked on 21 November 2026. How we check