Thinking Machines Lab releases Inkling 2 Mini, an Apache 2.0 looped model sized for 32GB laptops

Thinking Machines Lab releases Inkling 2 Mini, an Apache 2.0 looped model sized for 32GB laptops

中文版阅读

Thinking Machines Lab released Inkling 2 Mini on November 20, 2026, an open-weights multimodal model under the Apache 2.0 license designed for agentic coding and local laptop inference. Full checkpoints are downloadable on Hugging Face, day-one Mac support is live via Ollama, and hosted inference and fine-tuning run on Tinker.

Key architectural changes

  • Three-pass looped transformer: The 24-layer Mixture-of-Experts decoder routes 8 of 128 experts per token and loops its middle 12 layers three times with shared weights, yielding 48 effective layer passes without inflating parameter count.
  • Compact 20GB checkpoint: Total size is 26.4B parameters with 3.4B active per pass, rising to 5.4B effective active parameters across three loops. Expert layers train and run natively in 4-bit NVFP4 with 8-bit MXFP8 in the last four layers, avoiding post-training conversion loss after pruning from 384 experts and healing on 1T distilled plus 2.2T RL tokens.
  • 1M-token multimodal window: Accepts text and pixel images (40px to 4096px), outputs text, and supports up to 1 million tokens of context.
  • Day-one local and server runtimes: Runs offline on Apple silicon Macs with 32GB+ memory via Ollama or Exo Labs and on NVIDIA DGX Spark, plus SGLang, vLLM, Unsloth, NVIDIA Dynamo, and GGUF via llama.cpp; official materials do not yet publish Tinker API pricing.
  • Open variable-loop companion: The lab also released TML-d24-varloop-research (1 to 5 loops) on Hugging Face for interpretability and chain-of-thought monitoring research.

What it means for users

In evaluations published by Thinking Machines Lab, Inkling 2 Mini leads sub-100B open-weights peers on DeepSWE v1.1 (48.2% vs. 42.2% for Qwen3.8-27B), SWEBench Pro v2 Hard (77.9% vs. 73.8%), BrowseComp (79.0%), and DeepSearchQA (85.3%). Early Stanford Hazy Research and Scaling Intelligence Lab measurements on an NVIDIA DGX Spark (128GB, MXFP8) scored 0.703 intelligence per joule, ahead of Qwen3.6-35B-A3B (0.430) and Gemma-4-26B-A4B (0.159).

Teams testing local agents alongside hosted tools like Cursor or Claude Code should account for limitations documented in the model card. Looping deepens sequential reasoning without adding factual memory, so factual recall trails larger models and multi-turn or strict constraint tasks lag Qwen3.8-27B (50.0% MultiChallenge, 33.4% ComplexConstraints v1.1). While debugging code it can get stuck re-checking its work unless output tokens and agent turns are capped. Non-English output quality drops—Chinese runs showed clipped phrasing and mixed languages—minimum reasoning effort does not fully disable loops, and the lab recommends running agentic tools in a sandbox with explicit confirmation before file deletions.

Related tools

Download checkpoints or run local inference via Ollama and Hugging Face, or compare against the Qwen model family, Cursor, and Claude Code.

Sources: Thinking Machines Lab: Introducing Inkling 2 Mini · Inkling 2 Mini Model Card

Partager cet article