Inkling 2 Mini: can you run it locally, and is it any good?
Updated 2026-11-21. From Thinking Machines' launch post and model card; plan prices from the official pricing pages.
Yes, Inkling 2 Mini runs on a laptop, as long as it has 32 GB of memory or more. Thinking Machines released it on Nov 20, 2026 under Apache 2.0: 26.4B parameters, about 20 GB to download, 1M context, text and image input. On Thinking Machines' own benchmarks it edges out Qwen3.8 27B on agentic coding (48.2% vs 42.2% on DeepSWE) and ties it on the Artificial Analysis Index, but it trails Qwen on multi-turn chat and instructions with many constraints. The quickest start is ollama run inkling-2-mini.
Inkling 2 Mini: what we know and what we don't
| Question | Answer |
|---|---|
| Who made it | Thinking Machines Lab, the company behind Tinker and Inkling 2 Small |
| Released | Nov 20, 2026, free on Hugging Face |
| License | Apache 2.0, so commercial use is allowed |
| Size | 26.4B parameters, 3.4B active; the looped middle layers bring it to about 5.4B weights read per token |
| Download | About 20 GB for the official checkpoint (NVFP4 plus MXFP8); 17.6 GB for the 4-bit GGUF |
| Memory you need | 32 GB or more (Mac unified memory or GPU VRAM), per Thinking Machines |
| Context and input | Up to 1M tokens; reads text and images, writes text |
| Coding scores | 48.2% DeepSWE v1.1, 79.2% SWE-bench Verified (bash only) (vendor-reported) |
| Independent rankings | Not on LMArena's boards yet |
| Hosted API price | Available on Tinker, but we found no per-token price on the pages we checked |
From Thinking Machines' launch post and model card, checked 2026-11-21. Benchmark numbers are Thinking Machines' own, not independent tests.
What hardware Inkling 2 Mini needs
Plan on 32 GB. The model card lists Apple silicon Macs with 32 GB or more, one NVIDIA DGX Spark, or other GPUs with 32 GB of VRAM recommended. The official checkpoint is a single file of about 20 GB, already in 4-bit NVFP4 with the last four layers in MXFP8, so there is no separate quantizing step.
For llama.cpp and Llama App, the official GGUF builds are 17.6 GB (IQ4_NL, 4-bit) and 27.8 GB (Q8_0), plus a small image encoder file. The 1M context is the catch: long prompts need memory for the cache on top of the weights, so a 32 GB machine handles everyday coding sessions, not million-token documents.
- Works well: Macs with 48 GB or more (Thinking Machines demoed it offline on an M5 Pro MacBook Pro with 48 GB), a DGX Spark, or a 32 GB RTX 5090.
- Works: 32 GB Macs. A 24 GB GPU like the RTX 4090 should fit the 17.6 GB 4-bit build with a modest context window (our estimate; the card recommends 32 GB).
- No: 16 GB laptops and 8 GB to 16 GB GPUs. See the local model finder for models that fit.
Our per-quant, per-machine estimate is on the Inkling 2 Mini memory requirements page.
Why a 26B model punches above its size
Inkling 2 Mini is a looped transformer. Per the launch post, it has 24 layers, but the middle 12 run three times for every token, reusing the same weights. That gives it more thinking per token without more memory. It is also a mixture of experts, so only 3.4B parameters do work for each token, about 5.4B counting the loops.
That design is why it is quick on local hardware: a dense 27B model reads all 27B weights per token, while Inkling reads about a fifth of that. Thinking Machines is open about the limit: looping improves reasoning, not stored knowledge, so it knows about as many facts as other models its size and leans on search and tools to fill the gap.
Inkling 2 Mini vs Qwen3.8 27B
Qwen3.8 27B is the model to beat at this size, and it is the one Thinking Machines compares against most. On their numbers, Inkling comes out ahead on agentic coding and tool use, Qwen stays ahead on office automation, multi-turn chat and following long lists of constraints, and the two tie on the Artificial Analysis Index.
| Benchmark | Inkling 2 Mini | Qwen3.8 27B |
|---|---|---|
| DeepSWE v1.1 (agentic coding) | 48.2% | 42.2% |
| SWE-bench Verified (bash only) | 79.2% | 77.8% |
| Terminal Bench 2.1 | 71.6% | 73.0% |
| MCP Atlas (tool use) | 82.7% | 77.3% |
| AutomationBench-AA | 41.8% | 48.2% |
| Artificial Analysis Index v5.0 | 20.1 | 20.1 |
| MultiChallenge (multi-turn chat) | 50.0% | Higher (TML says Qwen leads) |
All scores from Thinking Machines' launch post (vendor-reported, setups vary by model). Hugging Face's team says that in their own vibe tests Inkling is "at least comparable with Qwen 3.8 27B" (Hugging Face blog).
The practical difference is speed and memory. Qwen3.8 27B is dense, so it reads every weight on every token; Inkling reads about 5.4B. On the same Mac, that makes Inkling several times faster by our estimates, with a smaller download. See Qwen3.8 27B requirements to compare.
Can it stand in for a paid coding plan?
As a fallback, yes. Run locally, it has no 5-hour or weekly cap, while Codex on ChatGPT Plus ($20) and Claude Code on Claude Pro ($20) both meter usage in 5-hour and weekly windows. Ollama can even launch Claude Code or OpenCode pointed at Inkling (ollama launch claude --model inkling-2-mini).
On quality, the paid plans still win. They run frontier models far above any 26B model, and Thinking Machines itself says Inkling trails larger models on hard reasoning. Use it for tests, bug fixes and small features when your plan hits its limit, or for code you can't send to the cloud. For how those windows work, see Codex limits explained and Claude Code limits explained.
How to run Inkling 2 Mini
- Ollama (Mac, Windows, Linux): install it, then run
ollama run inkling-2-mini. The Ollama listing is 20 GB with image input and a 1M context window. - Llama App or llama.cpp:
llama serve -hf ggml-org/Inkling-2-Mini-GGUF:IQ4_NLstarts a local OpenAI-compatible server with the 4-bit build, per the Hugging Face launch blog. - MLX on a Mac: mlx-lm supports it and can load the original checkpoint without re-quantizing.
- No hardware? Thinking Machines offers it on Tinker for inference and fine-tuning, and in the Tinker Playground for chat (launch post).
Comparing other open models first? The best local LLM for coding and best local LLM for a 32GB Mac pages rank the options per machine.
Sources
FAQ
What is Inkling 2 Mini?
It's a free, open-weights model from Thinking Machines Lab, released Nov 20, 2026 under Apache 2.0. It has 26.4B parameters, a 1M-token context and reads text and images. Its middle layers run three times per token, which Thinking Machines calls a looped transformer.
How much memory does Inkling 2 Mini need?
32 GB or more, per Thinking Machines. The official checkpoint is about 20 GB and the 4-bit GGUF is 17.6 GB; the rest goes to context and the system.
Can Inkling 2 Mini run on a 16GB Mac?
No. Even the 4-bit build is 17.6 GB before any context. Look at smaller models such as gpt-oss-20b or Gemma 4 12B in our local model finder.
Is Inkling 2 Mini better than Qwen3.8 27B?
For agentic coding and tool use, Thinking Machines' numbers say yes (48.2% vs 42.2% on DeepSWE v1.1). For multi-turn chat and complex instructions, Qwen3.8 27B is ahead. They tie on the Artificial Analysis Index at 20.1. All of these are vendor-reported.
How do I run Inkling 2 Mini with Ollama?
Install Ollama and run: ollama run inkling-2-mini. The download is about 20 GB. Ollama can also start Claude Code or OpenCode on it with ollama launch.
Is Inkling 2 Mini free for commercial use?
Yes. The weights are under Apache 2.0, which allows commercial use and changes.
Related
- Inkling 2 Mini memory requirements
- Local model finder
- Best local LLM for coding
- Best local LLM for a 32GB Mac
- Recursive Flint: the local coding model
- Qwen3.8 27B requirements
- Best local models
- Codex usage limit reached
- Claude Code limits explained
More guides
- Best local LLM for an RTX 5070 Ti
- MAI Code 2 in GitHub Copilot: price, plans and credits
- Claude Fable 5.2 and Mythos 5.2: prices, limits and what changed
- Claude Fable 5.2 vs Mythos 5.2: which one should you use?
- Is Claude Pro enough for Fable 5.2? Fable and Mythos limits by plan
- Claude Fable 5.2 in Claude Code: limits and cost
Not sure a plan beats the API for you? Try the plan vs API cost calculator.