Low-Refusal & Abliterated Open-Weight Model Catalog
Every model listed on bf16.ai passes continuous 6-vector behavioral verification across math reasoning, 1M context recall, strict JSON tool calling, and low-refusal fidelity.
Pass model="bf16/auto" to route automatically across healthy verified endpoints sorted by live TTFT and optional X-BF16-Max-Price-Per-M-Output ceiling.
Uncensored DeepSeek V4.1 Flash (Abliterated)
High-speed (~155 tok/s) multimodal (text + vision) weight-abliterated DeepSeek V4.1 Flash with 1M context window, deterministic tool calling, and zero preachy mid-workflow refusals — built for 24/7 MausBot, OpenClaw, and terminal coding agents.
Uncensored GLM-5.3 Flash Vision (Abliterated)
Multimodal (text + vision) orthogonal refusal-removed GLM-5.3 Flash with 1M context window, ~120 tok/s throughput, and strict JSON schema compliance for Browser-Use UI agents and DevSecOps pipelines.
Uncensored GLM-5.3 Flagship (Abliterated)
Full-size GLM-5.3 abliterated flagship model (~85 tok/s, 1M context) engineered for frontier coding, deep security vulnerability analysis, and multi-step Claude Code & Codex sessions.