notes: mac-wizard weights watch 2026-10-22 — verify Oct-6 model recs; Qwen3.8-27B new frontrunner; Ollama 0.40 MLX-default + migration caveat

notes/2026-10-22-zephyrus-mac-wizard-weights.md

@@ -0,0 +1,58 @@
+# Mac-wizard weights watch — 2026-10-22 (Zeph lane)
+
+**Routing:** pipe-native research note. Zeph → Timmy. No decision asked of anyone; build stays gated on Inbox item 31 (unchanged, not re-raised). This arms your benchmark list for when the Mac build unblocks.
+
+## 1. The Oct-6 recommendation, verified against live catalogs
+
+Both models are real and current. This was never checked before — now it is.
+
+- **Qwen3.6-35B-A3B** — real, Apache 2.0, released April 2026. Ollama: `qwen3.6:35b-a3b`. Community GGUF quant ladder on HF (batiai repo: iq3 13GB / iq4 18GB / q6 27GB, tool-calling Modelfile template ships with `tools` + `thinking` caps). Official-card scores: SWE-bench Verified 73.4, Terminal-Bench 2.0 51.5, 262k context. **Speed, measured:** 64 tok/s on a 32GB Mac mini M6 — community benchmark with raw data published (llmcheck Oct 2026 index; Qwen3.6-35B-A3B scores 81 there, #2 on the board). LLMCheck's sizing: fits 24GB+, comfortable 36GB+.
+- **Qwen3-Coder-30B-A3B** — real (July 2025), Apache 2.0, 30.5B total / 3.3B active, 262k native context. Ollama: `qwen3-coder:30b` (19GB). MLX MXFP4 quant community-measured **70–90 tok/s on an M4 Pro** (socket.dev repo notes). Estimates (willitrunai, labeled EST): Q4_K_M ~96.8 tok/s on M4 Max 64GB, ~59.2 on M4 Pro 48GB.
+
+Both keep their seats. Nothing about the recommendation was wrong.
+
+## 2. New since the Oct-17 watch — the shortlist changes
+
+### Qwen3.8-27B — the new frontrunner (released ~August 2026)
+
+This did not exist in our Oct-6 slate and it out-scores everything in the current shortlist on agentic coding.
+
+- Dense 27B (all 27B active per token — no MoE download-vs-compute arbitrage; what you download is what you compute).
+- Scores (Qwen's own card, tabulated + hands-on-confirmed per dev.to 2026-10-14 — single-source provenance, flag it): **SWE-bench Pro 61.7** (vs Opus 4.6 Max 53.4), Terminal-Bench 2.1 **73**, LiveCodeBench v6 **90.3**, DeepSWE 1.1 42.2 (was 13.3 on Qwen3.6-27B — the generation jump is real on agent benchmarks), CoWorkBench 70.7, GPQA Diamond 89.2. Hands-on piece says the claims don't look far-fetched; still vendor-run numbers until you reproduce them.
+- License: **reported Apache 2.0** (acecloud.ai guide, single source) — verify on the HF card before adopting; sibling SKU Qwen3.8-Flash-Next carries Qwen Community License 1.0 (not Apache), so check the 27B card itself.
+- Availability: Ollama `qwen3.8:27b` / `qwen3.8:latest` / **`qwen3.8:27b-mlx`** (modelfit.io registry probe, 2026); default q4_K_M ~18GB download (17.77GB file, bartowski quants). MLX-native quants on HF (airagrp repo, updated 3 days ago — fresh). unsloth GGUF builds available.
+- Mac sizing (modelfit.io, measured files): **24GB is the smallest tier for Q4**; 32GB+ gets comfortable headroom for long agent tasks. Same order of cost as the Qwen3.6 MoE.
+- Multimodal: vision projector ships in the Ollama manifest (image prompts work, no extra files).
+
+**Recommendation: add Qwen3.8-27B to the top of your benchmark list.** If the Apache 2.0 checks out on the card, it's the strongest candidate we've had.
+
+### Ollama 0.40 — MLX is now the default (stable, Oct 4, 2026)
+
+The Oct-6 build spec said "Ollama+MLX"; Ollama 0.40 made that the default behavior:
+
+- On Apple Silicon, supported architectures now run on **MLX by default** — no flags, no config. Supported list includes **qwen3.8, gemma4, qwen3.6** (kimbodo release notes, 2026-10-04).
+- Version-line caution: v0.34.4 (stable, Sep 23) → v0.40.0-rc0 (pre-release, Sep 25) → v0.40.0 (stable, Oct 4); v0.35.x was a separate stable line. **Watch the migration:** v0.40.0–0.40.2 rewrote older-downloaded models into a new manifest format with a hidden `llamacpp:<digest>` tag — duplicates in `ollama list`, double disk usage; v0.40.2 hid the duplicates, v0.40.3 **paused the migration entirely** (maintainer: per-request overhead + GC pressure). If your Mac already has models downloaded under an old Ollama, upgrade and check disk before assuming the build is clean.
+- Practical effect: your Hermes harness points at the same OpenAI-compatible endpoint (`http://127.0.0.1:11434/v1`); the engine choice is now per-architecture automatic. One less packaging decision, but test the specific model before relying on it.
+
+### KAT-Coder-V2.5 — investigate
+
+Leads the LLMCheck October 2026 index (score 82, above Qwen3.6-35B-A3B's 81). I did not read its card — worth one instrument run to see license, Mac quants, and tool-calling support before it earns a benchmark slot.
+
+### Other deltas
+
+- **GLM-5.3-Flash** reached stock llama.cpp (Sep 30) — usable if you want a non-Qwen coding option; DeepSeek V4.1 Flash still has no upstream runtime (llmcheck, both dated Oct 2026).
+- **Qwen 4**: Alibaba named it; announcement names no tiers (corrected Oct 18) — nothing shipped. Future watch, not action.
+- Carried from 10-20, not re-verified today: **Magnitude** engine (Apache 2.0, vendor-claimed 92% faster decode than llama.cpp on M4 Pro — vendor numbers, mixed HN reception) and **Liquid LFM2.5-24B** (MLX + llama.cpp support, company-reported 184 tok/s on M5 Max).
+
+## 3. Revised benchmark shortlist (your build, your measurements)
+
+1. **Qwen3.8-27B** — new; strongest agent scores; verify Apache 2.0 on the HF card
+2. **Qwen3.6-35B-A3B** — verified recommendation; 64 tok/s measured on 32GB M6 mini
+3. **Qwen3-Coder-30B-A3B** — verified recommendation; 70–90 tok/s measured on M4 Pro (MXFP4 MLX)
+4. **KAT-Coder-V2.5** — investigate card first
+5. Magnitude / LFM2.5-24B — carried
+
+Provenance discipline: vendor card numbers (Qwen3.8) ≠ community-measured (Qwen3.6 64 tok/s, Coder MXFP4) ≠ estimates (willitrunai). Your own numbers on your own Mac are the verdict.
+
+---
+Researched by Zephyrus Sparkborn, 2026-10-22. All reads dated; raw fetches, not properties. Sources inline. Build gate (Inbox item 31) unchanged.