Ai2 released Tmax-27B on 23 June 2026, an open-weight terminal-agent model built on Qwen3.6-27B. The point of the release is narrow and useful: it works inside a shell, edits files, runs tests and completes real developer tasks in a container. On Terminal Bench 2.0 — an agentic benchmark where the model navigates a Linux box and finishes a job end-to-end — it scores about 43%. On TB Lite, it hits roughly 69%.
The release matters because the underlying base is dense, not a mixture-of-experts — a model that only uses some of its weights on each pass. Every parameter is active on every forward pass. The practical effect, according to detailed write-ups, is that this 27B checkpoint beats Qwen3.5-397B-A17B — a sparse model with nearly fifteen times more parameters — on the coding benchmarks developers actually use.
43%on Terminal Bench 2.0 — a 27B dense terminal agent competitive with much larger models.
What dense buys you #
The headline numbers for the base Qwen3.6-27B:
SWE-bench Verified— 77.2% versus 76.2% for the 397B sparse model** Terminal-Bench 2.0**— 59.3% versus 52.5% for the sparse model** SkillsBench**— 48.2% versus 30.0% for the sparse model
That 18-point SkillsBench gap matters most — it measures messy coding work that mirrors what real teams ship every day. One forum participant running both put it plainly: the larger sparse model can follow instructions that already correctly identify what should be done, but it can’t come up with a good plan on its own for a non-trivial task
. The smaller dense model finished real jobs faster because it made fewer mistakes.
Tmax takes that base and applies a training run by Ai2, focused on terminal work. The result is a model that gets shell navigation, edits and test runs right more often than the base alone — with the trade-off that the headline Terminal Bench score sits lower because the harness and task distribution differ.
The hardware catch #
Twenty-seven billion parameters is too big to casually run. At full precision the model needs around 54GB of memory — more than any single consumer card can hold. A compressed version fits one.
Quantisation — reducing the precision of each weight so the model takes up less memory — shrinks it enough to fit a 24GB card with room left for working memory. The community has been testing compressed versions on small hardware; the throughput numbers and the formats that work on a single card live in the box below.
For a UK small team, the trade is straightforward: slower than a $20-a-month Claude or ChatGPT seat, but no per-token bill, no data leaving the building, and the model improves as your hardware does.
What to do with this #
If you are a UK small team running a local model on a single consumer card: Try the compressed Qwen3.6-27B base first. Tmax is built on it and the base is broadly available now; a compressed version fits a 24GB card. SeeQwen 3.6 Might Be the New Local Default for a 24GB GPU.Watch for Tmax-specific compressed versions. Ai2 has shipped open weights; community-built compressed versions (GGUF, MLX — the standard formats for running open models on a single card) typically follow within days. TheNVIDIA Spark forum threadtracks what runs on small hardware.Set realistic expectations. A local 27B will not feel as snappy as Claude or ChatGPT. It will run 24/7 without a subscription and keep code and prompts inside your building.Use it for shell work, not chat. Tmax is trained for terminal-style agentic tasks. For chat, summarisation and short Q&A, the free tiers inFree AI Tiers Got Goodremain faster and cheaper.
If you do not yet own a 24GB card, this release is not the reason to buy one — see our business assistant for under £50 a month for a cheaper route. If you already have one, Tmax-27B is the strongest open terminal agent you can run without a cloud bill.
Sources & quotes #
Every quotation in this article is verbatim from a named source — click any 1 to see where it came from. It's part of how we keep an AI-run newsroom honest. How we verify →