- 26B total / 5.4B active MoE with looped layers, using 3.1B unique weights per token
- Scores 20 on the Intelligence Index, matching Qwen3.8 27B and Nanbeige4.3-33B with 5-6x fewer active parameters
- Apache 2.0 licensed with text and image input, 1M context, and runs locally on a MacBook or DGX Spark
AI CURATED HOT STORY · 11/20/2026
Thinking Machines Open-Sources 26B Recurrent Model Inkling 2 Mini
Thinking Machines Lab, in collaboration with Hugging Face, has open-sourced Inkling 2 Mini, a multimodal mixture-of-experts model with 26B total parameters but only 3-5B activated per token. It scores 20 on the Artificial Analysis Intelligence Index and supports a 1M token context with text and image inputs, running locally on a MacBook or DGX Spark.
The model's key innovation is a recurrent Transformer architecture: the middle 12 layers share weights and loop three times, each iteration maintaining its own KV cache and reading gated messages from the previous round, thereby boosting effective capacity without increasing parameter count. vLLM provides Day-0 support on NVIDIA Blackwell, Hopper, and DGX Spark, with NVFP4/MXFP8 mixed-precision weights and a built-in DSpark speculative decoder for further acceleration. According to Hacker News, it achieves 48.2% on DeepSWE v1.1 and 46.5% on Vals Index v2.2, claiming leadership among all open-weight models under 100B parameters.
The release marks the arrival of efficient recurrent architectures in the open-source community, offering new possibilities for local deployment and long-context applications, and is poised to drive edge AI and privacy-sensitive scenarios while pressuring conventional Transformer models of similar scale.
Hourly heat curve
35 fully observed hoursThe line compares only sources observed throughout; gap hours are interpolated to keep the trend continuous.
Coverage timeline
Total 6 reports · Latest first- Artificial AnalysisT1·78 pts
- Arena.ai (@arena)T1·62 pts
- 26B-parameter looped transformer designed to run on a laptop
- Scores 20.1 on AA Index, positioned for long-horizon agentic tasks
- Live in Agent Arena, Code Arena and Text Arena; leaderboard scores pending
- Artificial Analysis (@ArtificialAnlys)T1·78 pts
- 26B total and 5.4B active parameters, with three passes through looped layers and only 3.1B unique weights per token
- Scores 20 on the Artificial Analysis Intelligence Index, on the Pareto frontier for intelligence vs. active parameters
- Apache 2.0 licensed with text and image input plus a 1M-token context, runnable locally on a MacBook or DGX Spark
- Hacker News AIT2·74 pts
- Looped transformer reruns middle layers three times per token to boost reasoning efficiency
- 26.4B total / 3.4B active MoE that runs locally on a laptop
- Claims to lead open-weights models under 100B, hitting 48.2% on DeepSWE v1.1
- vLLM 官方博客T1·78 pts
- Middle 12 layers run three passes with shared weights, each with its own KV cache and gated messages from the previous pass
- 26B total and 3.5B active parameters, 1M context, and native DSpark speculative decoding
- vLLM delivered day-0 support on Blackwell, Hopper, and DGX Spark
- Hugging Face BlogT1·74 pts
- About 26B physical but only 3B active parameters per token, suited to local hardware
- Some layers are looped three times to raise effective capacity without more parameters
- 1M-token context with image and text input, accelerated by NVFP4/MXFP8 mixed precision and a DSpark drafter