What Is Inkling 2 Mini? Thinking Machines' Laptop Model
The short answer
Inkling 2 Mini is Thinking Machines Lab’s small open-weights model, released November 20, 2026 under Apache 2.0, built to run on a 32GB laptop. It is a looped Mixture-of-Experts model — 26.4B total parameters, 3.4B active — that runs its middle layers three times per token to reason better without writing more tokens. In the vendor’s benchmarks it leads open models under 100B parameters on agentic coding (DeepSWE v1.1 48.2%). Facts below come from Thinking Machines’ announcement, the Hugging Face model card and Ollama, read November 21, 2026.
Key facts
| Inkling 2 Mini | |
|---|---|
| Released | November 20, 2026 |
| Licence | Apache 2.0 (plus Thinking Machines’ acceptable use policy) |
| Architecture | 24-layer MoE, 8 of 128 experts per token; middle 12 layers run 3 times |
| Parameters | 26.4B total, 3.4B active, 5.4B effective active with looping |
| Precision / size | Trained and shipped in NVFP4 (last 4 layers MXFP8), ~20GB checkpoint |
| Context | Up to 1M tokens |
| Input / output | Text and images in, text out |
| Runs on | Mac with 32GB+ memory, DGX Spark, GPUs with 32GB+ VRAM recommended |
| Software | Ollama, llama.cpp (GGUF), Unsloth, vLLM, SGLang, TokenSpeed, NVIDIA Dynamo, Exo; Tinker API |
Why “looped” matters
A standard transformer passes each token through its layers once. Inkling 2 Mini passes it through the first block once, the two middle blocks three times with shared weights, and the last block once — 48 layer applications from 24 stored layers. Thinking Machines says looping the middle layers gave a 1.51–1.91× compute-equivalent gain in training, and that a non-looped model needed 140% more inference compute to match the looped one’s loss on long agentic traces. The trade-off: looping improves reasoning, not stored knowledge, so factual recall is in line with other ~26B models.
Thinking Machines also released a research variant, TML-d24-varloop-research, trained with one to five loops, to study whether looping hides reasoning from chain-of-thought monitors. Its early finding: reasoning got shorter with more loops but stayed legible.
Benchmarks (vendor-reported)
| Benchmark | Inkling 2 Mini | Qwen3.8-27B | Gemma 4 31B |
|---|---|---|---|
| DeepSWE v1.1 | 48.2% | 42.2% | 0.3% |
| SWE-bench Verified (bash only) | 79.2% | 77.8% | 46.1% |
| Terminal-Bench 2.1 | 71.6% | 73.0% | 43.4% |
| MCP Atlas | 82.7% | 77.3% | 54.2% |
| AA Intelligence Index v5.0 | 20.1 | 20.1 | 4.4 |
| AA-Omniscience (factuality) | -8.6 | -6.4 | -48.5 |
Evaluation setups vary by model; treat these as the vendor’s claims until independent runs appear. Thinking Machines also reports the lowest misbehaviour and destructive-action rates of the group on its internal safety evals.
How to try it
- Mac (32GB+):
ollama run inkling-2-mini. - NVIDIA GPU or DGX Spark: follow the vLLM or SGLang recipe linked from the model card; GGUF quantizations from Unsloth fit smaller cards at some quality cost.
- No hardware: the Tinker playground offers inference-only access; Tinker also supports fine-tuning.
Pair it with an open-source coding agent such as OpenCode for a fully local setup — see best open-source Claude Code alternatives. Choosing between local coding models? See best local coding model for 32GB. The bigger sibling is covered in what is Inkling 2 Small.
Last verified: November 21, 2026.