Tinfield 1

Terminal work and long-horizon software engineering are what Bad Theory Labs built Tinfield 1 for. It is an open-weight model, a fine-tune of Alibaba's Qwen3.8-Flash-Next, and the lab presents it as its flagship. No release date is given by the developer. On Hugging Face the full-precision BF16 repository dates from 21 September 2026, but the quantized Compact and Mini versions appeared earlier, on 18 and 19 September, which makes 18 September the earliest dated item.

Specifications

According to the model card the model has 177 billion parameters in all, with 6.6 billion active for each token, a context of 256K tokens (262,144 is the base model's native length) and BF16 weights, although Hugging Face counts about 180 billion stored parameters, as for the base model. Two quantized versions, 72 GB and 61 GB in size, should fit on one machine with 64 GB, but the developer says it has not run them through its benchmarks. Like the base model, the weights use the Qwen Community License 1.0.

Results the developer reports

The developer's own evaluation, run with the mini-swe-agent harness, gives 33.0 on Terminal-Bench 4.0, where the base model scores 29.0, and 62 on DeepSWE v1.1, against 58.7. It also claims second place among open-weight models on the Terminal-Bench board. A check of the official Terminal-Bench 4.0 leaderboard on 22 September 2026 did not find Tinfield 1, so the claimed placing comes from the developer's own comparisons with published scores, not from a leaderboard entry.

What changed on the lab's own site

As of 22 September 2026 the lab's site carried no page for Tinfield 1; its homepage presented Tinfield as a model family still to be built. By 1 October 2026 the homepage showed Tinfield 1 as the lab's open-weight flagship, offered through its Runtime service at per-token rates, and again gave no release date. Whoever adopts it should try it on their own workloads.