Original title: thinkingmachines/Inkling-2-Mini
Released by Thinking Machines Lab on 11 20, this is the smaller version following Inkling 2 Small. It accepts text and images as input and outputs text, targeting agents and tool calling, coding assistants, dialogue, and RAG. It features a 24-layer decoder, where the middle 12 layers share weights and loop three times (resulting in 48 layer computations per token). Each layer selects 8 experts out of 128, mixing local and global attention. Total parameters are 26.4B with 3.4B active; accounting for the looping, this equates to approximately 5.4B effective active parameters. The maximum context length is 1M tokens. The weights are a mixed-precision checkpoint of approximately 20GB in NVFP4 / MXFP8. Officially, local inference can be run on a single DGX Spark, an Apple silicon device with more than 32GB of memory, or a GPU with 32GB of VRAM.
Why it matters In the official comparison table, coding and agent performance matches or exceeds that of similarly sized models; the single checkpoint is about twenty GB, runnable on Macs and single GPUs, under an Apache license
Upvotes 221.5 · Likes 636 · Downloads 136
Reasons
- Official comparisons with Gemma 4 31B, Muse Glimmer, Qwen3.8-27B, and Nemotron 3.5 Super VL: SWE-bench Verified 79.2%, SWE-bench Pro v2 (Hard) 77.9%, BrowseComp 79.0%, MCP Atlas 82.7%, OmniDocBench 92.8%; AA Index 20.1 ties with Qwen3.8-27B
- In internal safety tests, the rates of misoperations, hidden violations, and destructive actions are the lowest among the five models; officially stated as their first looping model, with research into the impact of looping on chain-of-thought supervision
- Day-one support from Hugging Face, SGLang, vLLM, llama.cpp, Ollama, etc.; Unsloth has released a GGUF quantized version; also callable and fine-tunable via the Tinker platform API
Usage
- Weights: https://huggingface.co/thinkingmachines/Inkling-2-Mini;在线试用:https://tinker.thinkingmachines.ai/playground
- For local inference, deploy using the SGLang / vLLM / TokenSpeed recipes linked in the card; for the GGUF version, see unsloth/Inkling-2-Mini-GGUF, run with llama.cpp or Ollama (Apple silicon)
- Post-training available via Miles, SkyRL; Tinker users should install the tml-renderers package and refer to tinker-cookbook
Risks
- Officially noted: weaker on knowledge-based Q&A compared to Inkling 2 Small, more prone to confidently wrong answers; when debugging code, it may get stuck in long reasoning loops of self-checking, so limiting output tokens and agent turns is recommended
- When using tools, it may delete or overwrite files without confirmation; officially recommended to run in a sandbox and add confirmations for destructive actions; Chinese tests have shown instances of truncated sentences and mixed Chinese-English text
- Even the minimum inference setting cannot completely disable thinking; looping layers make per-token computation higher than what active parameters suggest, support for looping models in local runtime is still new, and speed varies by framework. Evaluations are self-reported by the official team and were not run for this entry