Tinfield 1

Not yet independently verified. Dated only by Hugging Face repository timestamps; no press coverage was found. The Terminal-Bench 4.0 score and placing are self-reported, Tinfield 1 was absent from the official leaderboard on 22 September 2026, and the lab's own site lists Tinfield as 'not built'. We will update this when it can be confirmed, and remove this note.

Tinfield 1 is an open-weight model from Bad Theory Labs for terminal work and long-horizon software engineering. It is a fine-tune of Alibaba's Qwen3.8-Flash-Next. The developer gives no release date. The main BF16 repository on Hugging Face was created on 21 September 2026, but the Compact and Mini quantised builds were created on 18 and 19 September, so the earliest dated artefact is 18 September.

The model card describes 177 billion parameters in total with 6.6 billion active per token, a 256K-token context (the base model's native length is 262,144 tokens) and BF16 weights; Hugging Face's own count of the stored weights is about 180 billion, the same as the base model. Two quantised builds, of 72 GB and 61 GB, are said to run on a single 64 GB machine, but the developer notes that they have not been evaluated on its benchmarks. The weights are released under the Qwen Community License 1.0, following the base model.

In its own evaluation with the mini-swe-agent harness, the developer reports 33.0 on Terminal-Bench 4.0 (base model 29.0) and 62 on DeepSWE v1.1 (base model 58.7), and says Tinfield 1 is the second open-weight model on the Terminal-Bench board. When the official Terminal-Bench 4.0 leaderboard was checked on 22 September 2026, Tinfield 1 was not listed, so these placings are the developer's comparisons against published scores rather than leaderboard entries.

Bad Theory Labs' website has no Tinfield 1 page; its homepage describes Tinfield as a proposed model family with the status 'not built'. A Solana token named Bad Theory Labs (BTL), listed on CoinGecko, gives badtheorylabs.com as its website; whether the lab issued it is not stated. Buyers should test the model on their own workloads.