Full llama-server config for local Qwen3.6-35B-A3B with MTP speculative decoding
haydendevs · x · 2026-09-22
The author shares a complete llama-server launch command for running Qwen3.6-35B-A3B-MTP-UD-Q4KXL locally:
- 131072 context, 32768 prediction tokens, -fitt 1536
- Flash Attention on, K/V and draft caches quantized to q80 to save VRAM
- --spec-type draft-mtp enables MTP speculative decoding with --spec-draft-n-max 2
- --ctxcp 64, --no-mmap, --mlock, --no-warmup for memory behavior tuning
- Chat template kwargs preservethinking:true to keep reasoning traces
- Sampling: temp 0.6, top-p 0.95, top-k 20, no repeat penalty
Serving on 0.0.0.0:8080 as an OpenAI-compatible API — a ready-to-copy recipe for local MoE deployment.
Related event: Qwen3.6-35B Runs Smoothly on a Single RTX 4070(3 posts)→
More from Infra
- Banks reportedly halting compute lending as AI credit crunch begins — citrini · 2026-09-22
- California signs seven bills making AI data centers pay their own utility costs — The Verge AI · 2026-09-22
- O'Reilly builds a working data vocabulary for the semantic era, from warehouses to ontologies — rseroter · 2026-09-22
- UK's most powerful government AI supercomputer cost £225m — same as one bridge — emax · 2026-09-22
- exe.dev wins over developers: SSH into root VMs, plus the underrated Shelley coding agent — davidcrawshaw · 2026-09-22
- NVIDIA hosts trilateral meeting as Artificial Analysis becomes Korea sovereign AI evaluator — ArtificialAnlys · 2026-09-22