The Muse Glimmer Field Kit — Exact Commands, the Honest RAM Line for Every Machine, the opencode Configs, and the Hosted-API Tracker

The companion to the video. On August 10, 2026 Meta shipped Muse Glimmer — a free, Apache-2.0, ~30-billion-parameter agentic model with a 1.8B vision encoder and 128K context, distilled from the Muse Spark teacher, built to run on a normal laptop. This kit is everything we verified on launch day. PAGE 1 — WHAT SHIPPED: the full spec card (params, vision encoder, context, license fine print, BF16 vs 4-bit sizes) plus the family twist — same family as Meta's paid Muse Code CLI, opposite deal. PAGE 2 — THE BENCHMARK TABLE exactly as Meta published it (clearly labeled Meta's numbers, not ours): Glimmer-30B vs Gemma4-31B vs Qwen3.6-27B on SWE-Bench Pro, SWE-Bench Verified, MCP Atlas, and DeepSearch QA — including the honesty line where Meta's own chart shows Glimmer LOSING SWE-Bench Verified to Qwen by 1.2 points, plus the DFlash speculative-decoding speedup claims. PAGE 3 — CAN YOUR MACHINE RUN IT: the RAM tier table. 16GB Mac: no (run gpt-oss 20B quantized or qwen-coder 7B instead). 24GB Mac: technically loads, practically freezes. 32GB+: comfortable. NVIDIA/AMD/Windows: builds 'coming days' per Meta, with AMD's day-one blog linked. PAGE 4 — THE 24GB FREEZE RECEIPT from our own M4 Pro: free memory 57% to pinned 0–3%, the sampler heartbeat stretching 2s → 11s → 30s → 62s, and the iogpu.wired_limit sysctl warning (do NOT run it casually — it froze our whole Mac). PAGE 5 — THE EXACT COMMANDS: the ollama ≥0.32.7 requirement (older versions fail the pull with a 412 — we hit it twice), the 21GB pull, and the full tag/variant table. PAGE 6 — THE OPENCODE CONFIGS: the exact local-Ollama config from our launch-day runs (open model + open harness + $0, verified with Wi-Fi off) AND a hosted-endpoint variant ready to paste the day a provider goes live. PAGE 7 — THE HOSTED-API TRACKER: our 12:51 UTC launch-day audit showing none of Meta's named partners (Together, Fireworks, OpenRouter) serving it yet — every probe listed. PAGE 8 — every primary-source link, clickable: Meta's launch blog, the HF model card, the Ollama library page, AMD's day-one blog, and the LM Studio page. Independent field kit from Hyperautomation Labs — not affiliated with Meta.

By submitting, you'll be subscribed to the Hyperautomation AI Report newsletter — one practical AI email per week. Unsubscribe anytime.

Free. No spam. Unsubscribe anytime.