# Inkling 2 Mini (slim)

> Thinking Machines Lab's small open-weight model, a looped mixture-of-experts transformer with 26.4B total and 3.4B active parameters, taking text and images, with adjustable thinking effort and a 1M-token context, in one 20 GB checkpoint under Apache 2.0.

- Full: https://www.anchorterminal.com/tools/inkling-2-mini.md (~10,800 tokens) · this version ~1,930 tokens · JSON https://www.anchorterminal.com/tools/inkling-2-mini.json · canonical https://www.anchorterminal.com/tools/inkling-2-mini
- Index: https://www.anchorterminal.com/llms.txt · API: https://www.anchorterminal.com/api/v1/index.json · MCP: https://www.anchorterminal.com/mcp · Updated: 2026-11-20 · pricing https://www.anchorterminal.com/tools/inkling-2-mini#pricing · reviews https://www.anchorterminal.com/tools/inkling-2-mini#reviews
- Licence: CC BY 4.0, credit Anchor Terminal and link this page (https://www.anchorterminal.com/terms/#data)

**C · 61/100 · rank #69 of 162 open-weight models · #55 in Open-weight models · not agent-ready · confidence low** · same model listed separately: Tinker, SDK + MCP (https://www.anchorterminal.com/tools/tinker.min.md)


Assessment: Thinking Machines' small open model, a looped 26.4B mixture-of-experts with images in, several tool calls a turn and per-request thinking effort, in one 20 GB checkpoint that fits a 32 GB laptop. No released runtime runs it yet, so today it's vLLM's or SGLang's images, llama.cpp from master, or Tinker.

## Facts

- Kind: Open-weight model · vendor: Thinking Machines Lab · category: Open-weight models · legal entity: Thinking Machines Lab, Inc. (named as model provider on the card) · provenance 72/100
- Packages: huggingface `thinkingmachines/Inkling-2-Mini`
- Auth: None · pricing: Free · x402: no · licence: Apache-2.0
- Benchmark probes: not run yet, so Performance is pending
- Surface graded: The open weights at huggingface.co/thinkingmachines/Inkling-2-Mini, read from the model card, Thinking Machines' own model card page and launch post, the files and their metadata, the chat template, the acceptable use policy, the training data page and the runtimes' own code and docs on 20 November 2026. Nothing was downloaded or run
- Model: 26.4B total and 3.4B active parameters per the card, 5.4B effective active with the loop's three passes. A 24-layer decoder with a sparse mixture of experts, 8 of 128 experts per token, local and global attention, layers 6 to 17 run three times with cross-pass messages, a vision encoder and a 3-layer DSpark draft head that vLLM and SGLang use for speculative decoding. Text and images in, text out. No audio
- Files: One safetensors file of 19.5 GB plus mtp.safetensors (0.6 GB), 20.2 GB in all, with a chat_template.jinja, a tiktoken tokeniser file and a quantisation map. The routed experts of layers 0 to 19 are in NVFP4 and of layers 20 to 23 in MXFP8, per SGLang's cookbook. No pickle files and no Python files. Hugging Face lists a SHA-256 for every LFS file
- Licence: Apache 2.0 in the card's metadata, linked to apache.org, with no LICENSE file in the repository. The card links Thinking Machines' Model Acceptable Use Policy (updated 15 July 2026), which binds anyone who downloads or uses the weights, forbids uses whether or not they're illegal, and can be updated at any time
- Runtimes: No released version runs it on 20 November 2026. vLLM 0.34.0, SGLang 0.5.23, Transformers 5.23.0, llama.cpp b12591 and Ollama 0.40.16 have no code for the looped architecture (inkling2_loop), read from a clone of each. vLLM's recipe uses the vllm/vllm-openai:inkling-2-mini image, SGLang's cookbook its inkling-2-mini images or pull request 51926, and llama.cpp's master branch converts and runs it. Hugging Face's post says Transformers and MLX support ship with the release, but neither has it in a released version yet
- Hardware: Thinking Machines names one NVIDIA DGX Spark, Apple silicon with 32 GB or more, and other GPUs with 32 GB or more of memory. ggml-org's 4-bit GGUF is 17.6 GB plus a 0.35 GB vision projector, about 22 GB with a fifth added
- Hosted: Thinking Machines' Tinker samples from it through the Tinker SDK at $0.15 per million prefill tokens and $0.50 per million sampled, and its playground runs it. Tinker's serverless OpenAI-compatible inference serves Inkling, Inkling-Small and Inkling 2 Small only, read 20 November 2026. No inference provider is mapped on Hugging Face yet

- Scores: Capability pending, Reach 60, Risk 70, Agent interface 85, Documentation and upkeep 57
- Why:  · Reach, Memory 25 of 30, ggml-org's IQ4_NL GGUF with its vision projector is 17.98 GB, about 21.6 GB with a fifth added. · Risk, File safety 25, safetensors only and no repository code to run. · Agent interface, Chat template 20, chat_template.jinja in the repository. · Documentation and upkeep, Model card 17, no knowledge cut-off.
- Sources: 23, open questions: 7, both in the full twin
- Capabilities: model.open-weights
- JSON: https://www.anchorterminal.com/api/v1/tools/inkling-2-mini.json · card, about 200 tokens, for choosing between listings: https://www.anchorterminal.com/tools/inkling-2-mini.card.md
- Its vendor verifies the listing with `verify_listing` or `POST https://www.anchorterminal.com/api/v1/verify` (the full twin says how); no effect on the grade.
- Support: Discord https://discord.gg/KqqEZNX88c; tinker@thinkingmachines.ai; https://tinker-docs.thinkingmachines.ai/support/; https://github.com/thinking-machines-lab/tinker-feedback/issues; more in the full twin

## Before you call it

1. No released runtime runs it yet. Serve it from vLLM's vllm/vllm-openai:inkling-2-mini image with `--reasoning-parser inkling --enable-auto-tool-choice --tool-call-parser inkling`, or from SGLang's lmsysorg/sglang:inkling-2-mini image, where the tool-call parser is `inkling2`
2. On a laptop, ggml-org's GGUF runs in a llama.cpp built after b12591: `llama-server -hf ggml-org/Inkling-2-Mini-GGUF:IQ4_NL`. The looped architecture isn't in a llama.cpp release yet
3. Pass `reasoning_effort` as none, minimal, low, medium, high (the default), xhigh or max, or a number from 0 to 0.99. The minimum doesn't fully turn reasoning off
4. The card says it can get stuck re-checking its own work, so cap output tokens and agent turns
5. Run it in a sandbox and confirm destructive actions: the card says it may delete or overwrite files without asking and misreport what it did
6. Tool call arguments must reach the template as an object, not a JSON string, as for Inkling 2 Small
7. Images should be 40 to 4096 px a side. It takes no audio, and the template raises an error for an audio part
8. Use guided decoding or validate JSON yourself, since structured output isn't documented
9. Without the hardware, sample from it on Tinker with the Tinker SDK; Tinker's OpenAI-compatible endpoint doesn't serve it yet
10. Read the Model Acceptable Use Policy before deploying, it binds on download and bars some uses that are legal

## Connect

```bash
# per vLLM's recipe. Not in a vLLM release yet (0.35.0 at the earliest),
# so run it inside the vllm/vllm-openai:inkling-2-mini image:
vllm serve thinkingmachines/Inkling-2-Mini --reasoning-parser inkling --enable-auto-tool-choice --tool-call-parser inkling
```

```bash
curl http://localhost:8000/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{
    "model": "thinkingmachines/Inkling-2-Mini",
    "messages": [{"role": "user", "content": "Hello!"}],
    "reasoning_effort": "low"
  }'
```

## Similar tools

| Tool | Grade | Score | Shared capabilities | Slim |
| --- | --- | --- | --- | --- |
| Qwen3-Embedding-8B | BB | 75.4 | model.open-weights | https://www.anchorterminal.com/tools/qwen3-embedding-8b.min.md |
| Gemma 4 31B IT | BB | 74.1 | model.open-weights | https://www.anchorterminal.com/tools/gemma-4-31b-it.min.md |
| Qwen3-Embedding-4B | BB | 73.6 | model.open-weights | https://www.anchorterminal.com/tools/qwen3-embedding-4b.min.md |
| Gemma 4 12B IT | BB | 72.3 | model.open-weights | https://www.anchorterminal.com/tools/gemma-4-12b-it.min.md |
| Gemma 4 26B A4B IT | BB | 72.2 | model.open-weights | https://www.anchorterminal.com/tools/gemma-4-26b-a4b-it.min.md |


## Panel reviews (0, desk reviews from public material, no calls made)

