Pace model manifest — every model role, decided and on disk
Pace model manifest — every model role, decided and on disk
Canonical inventory of all models Pace needs (goal set 2026-06-10: all roles decided with runnable artifacts by end of day). One row per role; this doc is the single source of truth — update it when any decision changes. Embedded/ANE ports are perf infrastructure, not blockers: interim runtime is LM Studio / posttrainllm serve, which Pace already uses.
| Role | Model | Artifact (on this machine) | Size | Status 2026-06-10 |
|---|---|---|---|---|
| Planner (runtime default) | qwen/qwen3-30b-a3b | LM Studio | 18.6 GB | ON DISK, wired (Pace default) |
| Planner (specialist) | v11 DoRA on Qwen3-0.6B, served int8 g64 | ~/.cache/posttrainllm/runs/pace-planner-v11/baked-hf |
~0.5 GB eff. | TRAINING — ship-gate verdict tonight; ships only if it beats the gate, else 30B stays |
| Planner (ANE perf variant) | M8 28-block chain, int8 per-block | ~/.cache/posttrainllm/ane-w8b/ |
476 MB | ON DISK, numerics gate PASS, 17–22 tok/s |
| Screen VLM | A/B tonight: UI-Venus-1.5-2B-6bit vs Qwen3-VL-2B-Instruct-6bit (UI-Venus-8B as third column) | all three in LM Studio | 2.2–5.8 GB | DECIDING tonight (scripts/archive/eval_pace_vlm_ab.py, 12 fixtures, FakePace baseline 5/12); M4 Swift port (#266) follows the winner |
| ASR (voice) | WhisperKit openai_whisper-large-v3-turbo | ~/Documents/huggingface/models/argmaxinc/whisperkit-coreml/openai_whisper-large-v3-v20240930_turbo |
1.5 GB | ON DISK, qualified 2026-06-08 (9× realtime on ANE, perfect “Pace” accuracy); app integration = #292 (Swift work, not a model gap) |
| TTS | Apple AVSpeechSynthesizer (system) | none — OS framework (LocalTTSClient.swift) |
0 | DONE — no model needed, zero-cloud by construction |
| Embeddings (RAG) | Qwen3-Embedding-0.6B (decided 2026-06-09, replaced mxbai) | ~/.cache/posttrainllm/models/qwen3-embedding-0.6b/ |
1.2 GB | ON DISK; runtime retrieval is currently Spotlight-first (PaceSpotlightRetrievalConnector), embedding RAG is a stub contract (PaceLocalRetrieval.swift) — model ready when wiring lands |
| Dictation post-proc | Stage A rule-based (no model) | in pace repo, 26 tests | 0 | SHIPPED; Stage B model is conditional on a v11 BFCL trigger, deliberately not built |
Disk budget
On-disk total for the shipping set (30B planner + VLM winner + ASR + embeddings): ~22–24 GB, dominated by the 30B. If v11 ships and replaces the 30B for planning, the set drops to ~5–6 GB. The specialist question is therefore also a 4× bundle-size question.
What “done today” means per role
- Planner: v11 verdict lands tonight → either v11 ships (specialist) or the 30B is confirmed as the frozen decision. No third option; planner freezes either way.
- VLM: A/B runs right after the pipeline frees LM Studio → winner locked for #266.
- ASR / TTS / embeddings / dictation: already decided + on disk (audited today). Remaining work on these is app wiring, not model selection.
Stale items this manifest supersedes
- #299 (mxbai CoreML conversion) — obsolete since the Qwen3-Embedding swap.
- “Pace runtime planner = our specialist” assumptions anywhere — the default is the 30B until v11 proves otherwise on the gate.
Related
docs/prds/pace-planner-v11-ship-gate.md— tonight’s planner verdictdocs/prds/vlm-ab-uivenus-vs-qwen3vl.md— tonight’s VLM decisiondocs/prds/pace-task-loop-v1.md— first consumer of the full set- Memory:
whisperkit-qualified-2026-06-08,embedding-swap-2026-06-09,serve-quantize-2026-06-10,m8-int8-shipped-2026-06-10