177B total / 6.6B active per token (Sparse MoE; approximately 180B stored parameters)
256K
Qwen Community License 1.0
Open-weight agentic terminal and long-horizon software-engineering model built on Qwen3.8-Flash-Next. Tinfield 1 retains the base model's sparse multimodal architecture while specializing its behavior for terminal work, repository-scale coding, and extended software-engineering trajectories. It activates approximately 6.6B parameters per token from 177B total parameters; Hugging Face metadata counts approximately 180B stored parameters across the complete checkpoint. The release provides roughly 360 GB of BF16 safetensors and a 256K context window. Bad Theory Labs evaluated Tinfield 1 with mini-swe-agent at k=5 on the complete benchmark task sets, reporting 33.0 on the 66-task Terminal-Bench 4.0 and 62.0 on the 113-task DeepSWE v1.1, compared with 29.0 and 58.7 respectively for the underlying Qwen3.8-Flash-Next model. The developer describes Tinfield 1 as the second-highest-scoring open-weight model on the contemporary Terminal-Bench leaderboard and reports results above several frontier closed models, although those leaderboard comparisons use differing model harnesses. Official local-inference derivatives include the 72 GB Tinfield 1 Compact, using IQ2_XXS gate and up projections with IQ4_NL down projections, and the 61 GB Tinfield 1 Mini, using IQ2_XXS gate and up projections with range-searched Q2_0 down projections. Bad Theory Labs states that both compressed builds can run on a single 64 GB machine, but notes that their published benchmark scores were measured using the BF16 checkpoint and that the quantized versions have not yet been evaluated on those benchmarks. An official Q8_0 GGUF build is also available.