Views
No views yet
mlx-community/Qwen3.6-35B-A3B-4bit fine-tuned with our verifier-filtered LoRA (the same repair / tool-discipline / strict-JSON / domain data that trained the Hy3 models), fused to a standalone 18 GB MLX directory.qwen3_5_moe is natively supported).reasoning_content) and the answer to content. Toggle thinking off in LM Studio for direct answers, or give enough token budget for it to finish reasoning first.max_tokens too low fails silently, not loudly. If reasoning consumes the whole budget, content comes back empty with no error — this looks like a successful call that found nothing, not a truncation. Confirmed in a real production pipeline (859-row wiki-extraction run) that a max_tokens: 1600 cap returned zero results for every input before anyone noticed the cause; raising to 7000–9000 fixed it. If you're getting suspiciously empty/all-zero results with thinking left on, raise max_tokens before concluding the task legitimately had nothing to extract.1mlx_lm.chat --model <this-directory>
2mlx_lm.server --model <this-directory> --port 8080 # OpenAI /v1enable_thinking=False as a DIRECT kwarg to apply_chat_template (the nested form and /no_think are ignored).repeat_penalty: 1.0 explicitly in the request — don't trust a server's UI/global default. A nonzero repeat penalty punishes exactly the repeated structural tokens JSON needs ("name":, "type":, array delimiters), and can silently corrupt output on batch/array-heavy extraction tasks. Confirmed in production: a serving UI had repeat_penalty: 1.1 active globally, invisible to the request itself, on a strict-JSON extraction workload. A pipeline's full sampling contract (temperature, top-k, repeat-penalty) belongs in every request body, never assumed from server state — server defaults and UI settings can change between calls or restarts without you noticing.eval/receipts/sibling_qwen35b_stress_grade.json)