6,307 12 hours ago

Thinking Machines' small open-weights model that can run on a Mac: 26B parameters, 5.4B active, with looped layers for more computation per token.

vision tools thinking
ollama run inkling-2-mini

Details

12 hours ago

d183b1321cb5 · 20GB

{"architectures":["Inkling2LoopForConditionalGeneration"],"eos_token_id":200006,"model_type":"inklin
{ "do_sample": true, "temperature": 1.0, "eos_token_id": 200006, "pad_token_id": 199999 }
{ "model": { "parameter_groups": [ { "compute_dtype": "float32", "output_dtype": "float32", "require
{ "image_processor": { "do_convert_rgb": true, "do_normalize": true, "do_rescale": true, "image_mean
{ "additional_special_tokens": [ "<|endoftext|>", "<|message_user|>", "<|message_model|>", "<|messag
{ "version": "1.0", "truncation": null, "padding": null, "added_tokens": [ { "id": 199998, "content"
{ "tokenizer_class": "PreTrainedTokenizerFast", "model_max_length": 1048576, "clean_up_tokenization_
{"hidden_size": 2304, "num_hidden_layers": 3, "num_attention_heads": 24, "num_key_value_heads": 8, "
Apache License Version 2.0, January 2004 http://www.apache.org/licenses/ TERMS AND CONDITIONS FOR US
1250 tensors

Readme

Inkling 2 Mini is a small open-weights model from Thinking Machines, small enough to run on a Mac with 32GB+ memory. It has 26B parameters with 5.4B active per token, and loops its middle layers so it does more computation on each token instead of writing more tokens. It’s built for coding, agentic workflows such as computer use, design work that relies on strong visual reasoning, and finance tasks.

Get started

ollama run inkling-2-mini

Key features

  • Runs locally: 26B parameters with 5.4B active per token; runs on a Mac with 32GB+ memory or a DGX Spark.
  • Looping: 12 of its 24 layers run three times per token, so it does more computation on each token instead of writing more tokens.
  • Built for real work: coding, agentic workflows such as finance, computer use, and design with strong visual reasoning.
  • Native 4-bit: trained in NVFP4 and shipped in the same precision, so there’s no quantization loss.
  • Open weights: Apache 2.0.

Benchmarks

Inkling 2 Mini scores high on coding (79.2% on SWE-Bench Verified, 48.2% on DeepSWE), tool use for agents (82.7% on MCP Atlas) and deep research (85.3% on DeepSearchQA). image.png

Inkling 2 Mini also has strong visual reasoning for computer use and design: 76.5% on ScreenSpot Pro (finding things on a screen) and 92.8% on OmniDocBench (reading documents). image.png

Reference

Thinking Machines Blog

Thinking Machines Model Card