Inkling 2 Mini: open 26B looped model
Confirmed
The takeaway
On 20 Nov 2026 Thinking Machines Lab released Inkling 2 Mini, a 26.4B-parameter (3.4B active) mixture-of-experts “looped” transformer that runs its middle layers three times per token and fits in about 20 GB, with Apache-2.0 weights on Hugging Face.
Status
- Claim
Confirmed
- Our reporting
- High confidence
- Importance
- 3 of 5
- Last verified
- 22 November 2026
Your AI and this story
- GPT-6.2 Astra173 days after its cutoff
- Claude Opus 5.5143 days after its cutoff
- Gemini 4 Hydrogen234 days after its cutoff
- Grok 4.7173 days after its cutoff
None of these four assistants can know about it. The closest, Claude Opus 5.5, stops 143 days before it.
Key facts
- Size: 26.4B total, 3.4B active per token; 24 stored layers in four six-layer blocks, the two middle blocks run three times (48 layer applications per token); mostly NVFP4 weights, ~20 GB
- Benchmarks (lab): DeepSWE v1.1 48.2%, Vals Index v2.2 46.5% (best open model under 100B), AA Index v5.0 20.1, SWE-bench Verified 79.2%, BrowseComp 79.0%; trails Qwen3.8-27B on ComplexConstraints (33.4%)
- Artificial Analysis (Nov 20, via Techmeme): matches Qwen3.8 27B (xhigh) on its Intelligence Index with about 5x fewer active parameters
- Availability: open weights, Apache-2.0 (huggingface.co/thinkingmachines/Inkling-2-Mini); Tinker; SGLang, vLLM, Ollama, llama.cpp (GGUF)
- Research model TML-d24-varloop-research (1–5 passes): no-CoT time horizon rose from 16 s at one pass to 37 s at three
- Mira Murati on X (506k views): ‘At this scale, its thought process remains monitorable; this may not stay the case’; the lab reports no meaningful loss of reasoning monitorability
- ARC Prize verified scores (20 Nov): ARC-AGI-2 49.3% at $0.14 per task, ARC-AGI-1 84.5% at $0.06; 47% cheaper per task than Inkling 2 Small on ARC-AGI-2 but 32 points lower
What happened
Two weeks after Inkling 2 Small (594B), Thinking Machines released a small sibling for local use. Instead of more layers it reuses the middle of the network: each token passes through the two central blocks three times, which the lab says lets the model “think more and write less”.
Why it matters
Looping adds hidden computation that does not appear as visible chain-of-thought. The lab says reasoning is still monitorable at this size, and released a variable-loop model so others can study the point at which that stops being true.
Sources
5 sources from 3 sites. Numbers match the chips in the text.
5 sources: 4 primary, 1 reaction
Primary
- Thinking Machines: Inkling 2 Minithinkingmachines.ai, official
- Hugging Face: thinkingmachines/Inkling-2-Minihuggingface.co, code
- Hugging Face blog: Welcome Inkling 2 Minihuggingface.co, docs
- Mira Murati on Xx.com, official
Reactions
- ARC Prize on X: Inkling 2 Mini on ARC-AGI (Verified)x.com, discussion
Changes
- ARC Prize verified scores (news agent)
- Filed (news agent)