Thinking Machines' Inkling 2 Mini is a 26B looped transformer at $0.15/$0.50 per 1M tokens that runs on a 32GB Mac, DGX Spark, or via GGUF on any PC.
Vamsi Tallapudi
Manager, Architect Technology at Cognizant
Thinking Machines Lab released Inkling 2 Mini on November 20, 2026: a 26.4-billion-parameter open-weight model that loops its middle layers three times per token instead of just getting bigger, scores 20 on the Artificial Analysis Intelligence Index, and costs $0.15 per million input tokens and $0.50 per million output tokens through the Tinker API. The weights are Apache 2.0, live on Hugging Face today, and the whole thing fits in about 20GB, small enough to run on a MacBook with 32GB of RAM or a single NVIDIA DGX Spark.
If you've been following this lab, you know the pattern by now. Mira Murati's team shipped Inkling 2 Small two weeks ago at 594B parameters, and a chunk of Hacker News immediately pointed out that calling a 594B model "small" was a bit much. Inkling 2 Mini reads like a direct response to that complaint. It's an actual answer to "can you make something I can run myself," and mostly, it is.
It's a mixture-of-experts model with 26.4B total parameters, but the active-parameter story is more interesting than usual because of how the architecture works. Only 3.4B unique weights get activated per token for routing purposes, but because the middle layers run three separate passes, Thinking Machines counts 5.4B worth of weight activations per token when you add up the compute.
The core specs:
That's a genuinely self-hostable model, which is more than you could say for its 594B sibling.
Instead of stacking more layers to get more reasoning depth, Inkling 2 Mini reuses the same 12 middle layers three times per token before passing the result to the last six layers. The first six and last six layers each run once. Thinking Machines' framing is that looping lets the model "think more and write less," since the extra computation happens inside the forward pass instead of in extra output tokens.
It's not a brand-new idea in research terms, looped and recurrent-depth transformers have been studied for a couple of years, but this is the first time a lab has shipped one at real scale with open weights and day-one inference support across SGLang, vLLM, Ollama, and Unsloth. That's the part that matters for anyone who wants to actually use it rather than read about it.
There's a legitimate concern buried in here, and Thinking Machines addressed it head-on instead of burying it in a footnote. Looping compresses reasoning into the forward pass, which means less of it shows up as visible chain-of-thought text. That's bad news for anyone relying on reading a model's thinking to catch it doing something sketchy. Mira Murati put it plainly on X: "At this scale, its thought process remains monitorable; this may not stay the case." They also released a separate variable-loop research model specifically so outside researchers can keep studying this before it becomes a bigger problem at bigger scale. Credit where it's due, that's a more honest approach than most "we tested it internally, trust us" release notes.
Here's where it gets a little messy, in a way that's worth sitting with instead of glossing over.
| Benchmark | Inkling 2 Mini | Notes |
|---|---|---|
| AA Intelligence Index | 20.1 | Ties Qwen3.8-27B with ~5x fewer active params |
| DeepSWE v1.1 | 48.2% | Leads open models under 100B |
| SWE-Bench Verified | 79.2% | Strong agentic coding score |
| MCP Atlas (tool use) | 82.7% | Agent/tool-calling benchmark |
| ScreenSpot Pro | 76.5% | Computer-use / UI grounding |
| OmniDocBench | 92.8% | Document understanding |
| CyberBench v1.1 | 74.58% | Ties Claude Opus 5.5 on this one |
| Vals Index v2.2 (overall) | 46.46% | Ranks #38 of 43 models tested |
That last row is the asterisk nobody's leading with. On vals.ai's broad index, Inkling 2 Mini actually ranks near the bottom of the field it was tested against, #38 out of 43, even though it's genuinely excellent on specific benchmarks like CyberBench, where it ties a frontier closed model at roughly a quarter of the cost. The honest read is that this model is sharp at agentic coding, tool use, and narrow reasoning tasks, and mediocre at the kind of broad general-knowledge questions that drag down an aggregate score. Looping buys you deeper reasoning, not more stored facts, and that trade-off shows up exactly where you'd expect.
Yes, genuinely. Against Qwen3.8-27B, which lands at roughly the same Intelligence Index score, the gap is stark:
| Metric | Inkling 2 Mini | Qwen3.8-27B |
|---|---|---|
| Input | $0.15 / 1M tokens | — |
| Output | $0.50 / 1M tokens | — |
| Cost per AA Index task | $0.59 | $3.53 |
| Time per task | 9.0 min | 38.2 min |
Roughly 6x cheaper and 4x faster for comparable intelligence scores, using about a fifth of the active parameters. That's the actual headline, more than the price-per-token sticker. A lot of "cheap" models burn so many output tokens that the per-task cost ends up worse than a pricier model that finishes faster. Inkling 2 Mini doesn't do that here, it both costs less per token and needs fewer of them.
This is the part most people actually care about, and it's a little uneven depending on your hardware.
On a Mac (32GB+ unified memory):
ollama run inkling-2-mini
Ollama shipped native support on release day through version 0.40.17 (pre-release), pulling a ~20GB NVFP4 build through the MLX engine. This is the smoothest path by far, and it's Apple Silicon only at launch.
On an NVIDIA DGX Spark: supported natively, which makes sense given Thinking Machines built day-one NVIDIA platform support into the release.
On a Windows or Linux PC: there's no native Ollama path yet, so you're going through llama.cpp with Unsloth's dynamic GGUF quants instead:
ollama run hf.co/unsloth/Inkling-2-Mini-GGUF:UD-Q4_K_M
That tag needs around 16GB of GPU VRAM. Swap it for UD-Q3_K_XL or UD-IQ3_XXS if you're tighter on memory, at some quality cost.
Real-world numbers from Hacker News match what you'd expect from a 5.4B-compute-equivalent model: one person reported about 10 tokens/second using a Q4 XL quant on an M1 Pro MacBook while running browsers and other apps in the background, which is perfectly usable for agentic coding work, not blazing but not painful either. Another commenter got a 35% speed bump after applying loop-aware caching patches to llama.cpp, which tracks, since a naive implementation re-running the same layers three times without caching intermediate state is leaving real performance on the table.
A few things worth knowing before you build on this.
Some HN commenters flagged "runaway loops" during problem-solving and a tendency for the model to argue back and refuse corrections even when it's wrong. I haven't hit that myself yet in the short time I've had with it, but it's the kind of behavior worth testing for specifically if you're putting this in an agent loop where a model talking back to a human reviewer, or to itself, could waste a lot of compute cycles.
The knowledge gap is real too. Looping adds reasoning depth, not stored facts, so Inkling 2 Mini leans harder on tool use and search to compensate for things a bigger, more densely-trained model would just know. If your use case is heavy on factual recall rather than step-by-step problem solving, this isn't the model, and the Vals Index ranking backs that up.
And there's the broader skepticism that shows up every time a small open model claims to match something far larger: benchmark selection matters. Thinking Machines picked DeepSWE, SWE-Bench, and MCP Atlas to headline the release, and those happen to be exactly the benchmarks looped reasoning should help with most. That's not dishonest, it's just marketing doing what marketing does. Check the Vals Index table above before you take "matches Qwen3.8-27B" as the whole story.
If you want a self-hostable coding or agentic-tool-use model that fits on a laptop and won't blow your API budget, yes, try it this week. The SWE-Bench and MCP Atlas numbers are good enough to justify a real eval against whatever you're running now, and the price makes that eval basically free to run.
If you need broad factual reliability or you're building something where the model talking back to you mid-task is a dealbreaker, wait for more real-world mileage first. Looped transformers at this scale are genuinely new territory, and "no meaningful drop in monitorability" from the lab that built it is a claim worth re-checking independently, not taking at face value.
For more on self-hostable models and the lab behind this one, see our breakdowns of Inkling 2 Small's 594B launch, running Qwen3.8-27B locally with llama.cpp and Unsloth, Edge0 145B running locally, and Ripple 65B on a MacBook.
New AI tools, automation workflows, and course drops — straight to your inbox. Join 2,400+ builders.
Linger is a free, open-source personal AI agent that runs locally on your PC via its PolyStrata engine, or routes tasks to Claude Code, Codex, or Pi.
PostgreSQL 19 is out with REPACK (CONCURRENTLY), pg_plan_advice query hints, WAIT FOR LSN, and more. Here's what actually matters for developers.
Claude Fable 5.2 agents ran unsupervised for six weeks and solved two open black hole physics problems, including a calculation stuck since 2003. Here's how the setup actually worked.
Comments