QWEN3.6-35B-A3B-MTP + VISION: DOES --MMPROJ KILL THE MTP DRAFT?

I've seen maybe some problems in the past with this model, so maybe an update it is welcome to all thos ewho use it. The concern: https://www.reddit.com/r/llamacpp/comments... The answer: No — --mmproj does not kill the MTP draft. Not on the vision turn, and not for the rest of the session. llama.cpp b9620 · Qwen3.6-35B-A3B-MTP · M5 Max · unedited. Everything is on screen: the server log on the left, the model answering on the right. Watch 2:14 — the image description lands while the same log prints "draft acceptance = 0.61950" at 104.68 t/s. One frame, both facts. THE LAUNCH llama-server -hf unsloth/Qwen3.6-35B-A3B-MTP-GGUF:UD-Q4_K_XL \ --mmproj .../mmproj-BF16.gguf \ -ngl 99 -c 262144 -fa on -np 1 \ --spec-type draft-mtp --spec-draft-n-max 2 --port 8081 Both budgets appear side by side at 0:38 — the projector does not evict the draft context: [mtmd] estimated worst-case memory usage of mmproj is 1134.00 MiB [spec] estimated memory usage of MTP context is 826.70 MiB common_speculative_impl_draft_mtp: adding speculative implementation 'draft-mtp' CHAPTERS 0:00 CO_DE, opening the hajimi project 0:20 New desk → new terminal, side by side with the chat 0:30 Launching the MTP + mmproj server 0:38 Boot log: mmproj AND MTP draft both initialize 0:45 TEXT test — "quick audit of this repo's entry points" 1:14 Drafting on text: 0.862 → 0.955 acceptance 1:50 The audit result (real tables, local model) 2:00 VISION test — "describe this image" 2:02 process_mtmd: encoding mtmd batch 2:14 The answer + draft acceptance 0.6195 @ 104.68 t/s 2:20 Conclusion THE NUMBERS Text turns 0.862 / 0.955 / 0.828 / 0.739 acceptance · 79–100 t/s Vision turn 0.6195 (394 accepted / 636 generated) · 104.68 t/s — fastest eval of the whole session After vision 0.680 — a text turn AFTER the image still drafts Session total 1474 / 2020 draft tokens accepted (~73%) across 7 turns Acceptance does drop on the vision turn. That is the draft head being less sure about tokens that follow an image — lower acceptance, not a disabled draft. Throughput did not suffer. If you do OCR/grounding, the loader warns: "Qwen-VL models require at minimum 1024 image tokens … try --image-min-tokens 1024" THE WORKSPACE The app is CO_DE — a native macOS workspace where CLI coding agents (Claude Code, Codex, Cursor, Gemini) sit next to your terminals, with a local models in CLIs like PI, Opencode, or in native chat the orchestrator seat. That's why the server log and the model answering it are on one screen: same machine, same session, nothing spliced. CO_DE: https://github.com/gelubodrug/co_de Reddit thread: https://www.reddit.com/r/llamacpp/comments...