Short answer: Inkling 2 Mini is Thinking Machines Lab’s new open-weights coding and agent model, released under Apache 2.0, that fits in about 20GB and runs on a 32GB Mac, a DGX Spark or a 32GB GPU. It is a 26.4B parameter Mixture-of-Experts model that loops its middle layers three times per token, and Thinking Machines reports it leads every open model under 100B parameters on agentic coding tests such as DeepSWE v1.1 (48.2%). The caveat: those scores are self-reported, and the model trails Qwen3.8-27B on multi-turn chat, strict instruction following and factual recall.
Key points
- Open weights on Hugging Face under Apache 2.0, plus a research variant with a variable number of loops. Hosted inference and fine-tuning run on Thinking Machines’ Tinker platform, with free chat in the Tinker Playground.
- 26.4B total parameters, 3.4B active, and about 5.4B effective active per token once the three loop passes are counted. Each token is routed to 8 of 128 experts.
- About 20GB on disk, shipped in the same NVFP4 and MXFP8 precision it was trained in, so there is no extra quantization loss in the official checkpoint.
- Context window up to 1M tokens. Input is text and images, output is text.
- Runs today in Ollama, vLLM, SGLang, llama.cpp (Unsloth GGUF builds), NVIDIA Dynamo, Exo and Llama App.
- Reported coding scores: 48.2% on DeepSWE v1.1 versus 42.2% for Qwen3.8-27B, and 79.2% on SWE-bench Verified (bash only) versus 77.8%. Qwen3.8-27B still edges it on Terminal Bench v2.1, 73.0% to 71.6%.
What shipped with Inkling 2 Mini
Thinking Machines Lab released Inkling 2 Mini as a smaller sibling of Inkling 2 Small, sized for local hardware. The launch post reached the top of Hacker News with more than 1,000 points, and the Unsloth GGUF build passed 200,000 downloads on Hugging Face within about a day, which tells you who it was built for: developers who want a capable model on their own machine.
The headline idea is the looped transformer. The model has 24 layers split into four blocks. The first and last blocks run once, while the two middle blocks run three times with shared weights, giving 48 layer passes per token from 24 stored layers. Thinking Machines says this lets the model reason more before it writes, so it reaches answers with fewer output tokens. In its research variant, five passes cut average reasoning length on GPQA Diamond from 3,703 tokens to 1,661 while the chain of thought stayed readable.
The company also trained most of the model natively in 4-bit NVFP4, with the last four layers in 8-bit MXFP8, and pruned two thirds of the experts during training to shrink it. That is why the official checkpoint lands near 20GB instead of needing a separate community quantization to fit on a laptop.
How Inkling 2 Mini compares with other small open models
Thinking Machines compares it against Gemma 4 31B, Muse Glimmer, Qwen3.8-27B and Nemotron 3.5 Super VL. The real fight is with Qwen3.8-27B, which is dense, so all 27B parameters work on every token.
| Benchmark (as reported by Thinking Machines) | Inkling 2 Mini | Qwen3.8-27B |
|---|---|---|
| DeepSWE v1.1 | 48.2% | 42.2% |
| SWE-bench Verified, bash only | 79.2% | 77.8% |
| SWE-bench Pro v2 Hard | 77.9% | 73.8% |
| Terminal Bench v2.1 | 71.6% | 73.0% |
| MCP Atlas (tool use) | 82.7% | 77.3% |
| BrowseComp, with compaction | 79.0% | 61.6% |
| AA-Omniscience non-hallucination rate | 63.4% | 71.4% |
Two things stand out. First, it is strongest exactly where builders run local models: coding agents, tool calls through MCP and search-heavy tasks. Second, looping does not add knowledge. Thinking Machines says factual recall tracks total parameter count, and Inkling 2 Mini hallucinates more than Qwen3.8-27B on AA-Omniscience. It is also weaker on multi-turn chat (50.0% on MultiChallenge) and on prompts with many constraints (33.4% on ComplexConstraints v1.1). Hosted models are still well ahead on the hardest coding tests; for contrast, DeepSeek V4.1 Flash 1017 reports 77% on DeepSWE.
What Inkling 2 Mini means for builders
For most developers the practical change is simple: a model that competes with the best small open coder now runs offline on a 32GB MacBook, and it is free to use commercially under Apache 2.0. That matters if you want an agent on private code, a local fallback when an API rate limits you, or a cheap model for background agents that would burn tokens on a paid plan.
- Fastest way to try it:
ollama run inkling-2-mini. Ollama also lists launch commands for coding agents, for exampleollama launch claude --model inkling-2-miniandollama launch opencode --model inkling-2-mini. - For a server or GPU box: vLLM and SGLang have official recipes, and Unsloth publishes quantized GGUF files for llama.cpp if you have less than 32GB.
- For fine-tuning: Tinker offers hosted fine-tuning and inference, and SkyRL and Miles support RL post-training. Thinking Machines did not publish Tinker prices in the launch post, so check Tinker before you plan around it.
- Real-world speed: one Hacker News commenter reported about 10 tokens per second on an M1 Pro with 32GB using a 4-bit quant. That is a single user report, not a benchmark.
Running locally also keeps your code and prompts on your machine, which is worth more after reports like Windows 11 diagnostics flagging which coding agents you run. If you have been weighing other local options, compare it with Strata running Qwen3.8-Flash-Next on a 12GB gaming GPU for bigger models on smaller cards, or Meta’s Muse Spark Open weights if you have data center hardware.
Our take: Inkling 2 Mini is the most interesting local coding model of the moment because of the architecture, not only the scores. If looping keeps delivering this much reasoning per parameter, the gap between a laptop model and a paid API will keep shrinking for agent work. Treat it as a strong local agent and tool caller, and keep a hosted model for factual questions and long, picky chat.
What to watch next
- Independent results. The comparison table is Thinking Machines’ own, and the company notes evaluation setups vary by model.
- Speed on common hardware once Ollama, MLX and llama.cpp builds settle, especially on 32GB Macs where the model barely fits next to an IDE and browser.
- Whether other labs follow with looped open models. Thinking Machines is among the first to ship one openly, and it is working with interpretability researchers at NDIF and Goodfire to study them.
- Tinker pricing and any limits on hosted fine-tuning for Inkling 2 Mini.



