Inkling 2 Mini: A 26B Open Model for 32GB Laptops, vs Qwen
Thinking Machines released Inkling 2 Mini on Nov 20: 26.4B params, about 20GB, Apache 2.0, runs on 32GB Macs via Ollama. What it beats and where it trails.

Thinking Machines released Inkling 2 Mini on Nov 20: 26.4B params, about 20GB, Apache 2.0, runs on 32GB Macs via Ollama. What it beats and where it trails.

On November 20, 2026, Thinking Machines Lab published Introducing Inkling 2 Mini ↗, calling it "a fast, powerful model that's small enough to run on your laptop." The model card ↗ lists the license as Apache 2.0 and the release date as November 20. The weights are on Hugging Face ↗, where Hugging Face's own team posted a launch guide ↗ the same day. Unite.AI ↗ covered the release too. Thinking Machines on X ↗ framed it as an answer to complaints that Inkling 2 Small "was too big at 594B parameters."
The picture above is an original graphic created by aitoolranked. We have not run Inkling 2 Mini.
| Spec | Inkling 2 Mini |
|---|---|
| Total parameters | 26.4B |
| Active per token | 3.4B (5.4B with loops) |
| Layers | 24, middle 12 run 3 times |
| Experts | 8 of 128 per token |
| Context | Up to 1M tokens |
| Inputs | Text and images |
| Checkpoint | About 20GB, NVFP4 + MXFP8 |
The "looped" part is the new idea. The middle 12 layers share weights and run three times for every token, so 24 stored layers act like 48. Thinking Machines says this lets the model "apply more reasoning and reach the right answer with fewer tokens." It also ships a research variant, TML-d24-varloop-research, that can run a different number of passes.
The model card names three local targets: one NVIDIA DGX Spark, Apple silicon Macs with 32GB or more of memory, and other GPUs with 32GB or more of VRAM recommended. Ollama ↗ said it is available with version 0.40.17, then in pre-release, and the Ollama library page ↗ lists ollama run inkling-2-mini. Hugging Face's guide covers Transformers, llama.cpp, MLX, and SGLang, and links 4-bit and 8-bit GGUF files. Hosted access runs through Thinking Machines' Tinker platform and the Tinker Playground. The launch post names no price for Tinker use, so this post does not give one.
All scores below come from the launch post's table, and each row's test setup is the vendor's.
| Benchmark | Inkling 2 Mini | Qwen3.8-27B |
|---|---|---|
| AA Index v5.0 | 20.1 | 20.1 |
| DeepSWE v1.1 | 48.2% | 42.2% |
| SWE-bench Pro v2 Hard | 77.9% | 73.8% |
| SWE-bench Verified | 79.2% | 77.8% |
| Terminal Bench v2.1 | 71.6% | 73.0% |
| MCP Atlas | 82.7% | 77.3% |
The company is upfront about the gaps. It says Inkling 2 Mini "is not the strongest small model for multi-turn chat or for instructions with many constraints," with 50.0% on MultiChallenge and 33.4% on ComplexConstraints v1.1, behind Qwen3.8-27B on both. It also says looping "doesn't enable its weights to store more facts," so closed-book knowledge is in line with other models of its size.
One outside number exists. ARC Prize ↗ posted verified scores of 49.3% on ARC-AGI-2 at $0.14 per task and 84.5% on ARC-AGI-1 at $0.06 per task, and noted it "costs 47% less per task than Inkling 2 Small on ARC-AGI-2, but scored 32 percentage points lower."
"In our vibe tests, it's at least comparable with Qwen 3.8 27B." — Hugging Face's launch guide ↗, which adds that more validation is needed
"Small enough to run on a Mac with 32GB+ memory." — Ollama ↗ (@ollama)
"Looped models are weird. A layer can write to itself or read from "future" layers." — Goodfire ↗ (@GoodfireAI), which says it had early access
"~8 tok/s on CPU is pretty fast!" — RedFaster ↗ (@RedFasterX)
"6gb vram + 16gb ram can?" — Basyir Rionix ↗ (@rionix88), asking the question many readers will have
Reddit was not reachable for this story.
This is the Inkling you can actually run. Our Inkling 2 Small note said not to buy hardware for that 594B model. Mini flips that. An Apache 2.0 file of about 20GB that the vendor supports on a 32GB Mac is a real local option, not a lab toy.
Check your memory first. The card's floor is 32GB of unified memory or VRAM. On a 16GB laptop or a 6GB graphics card, plan on a hosted endpoint instead. The CPU speed figure above is one user's post, not a benchmark.
Pick it for agentic coding, not for chat. On the vendor's table it leads Qwen3.8-27B on coding and tool-use rows and ties it on the AA Index. It trails on multi-turn chat and constraint-heavy instructions, which is exactly what a writing assistant needs. Our local LLM comparison and open-source LLM roundup cover the alternatives, and our GGUF explainer helps you pick a quant. For coding agents, compare it with what our coding guide recommends before switching.
Wait for independent coding evals. ARC Prize is the only outside score so far, and Arena says its numbers are "coming soon." Vendor rows with vendor harnesses are a reason to try, not a reason to migrate.
Verdict: the most interesting small open model this week, with a permissive license and a realistic hardware bar. Try it locally if you have 32GB; keep your current model until third-party coding results land.