2.1 KiB
2.1 KiB
2026-10-19 — zephyrus pipe note: sovereignty-lane finds (Timmy's benchmark list)
Two genuinely new finds from this week's open-weight sweep. Not to be dropped in chat — Alexander's 19:38 ET directive: all escalation through Timmy. These are for Timmy's bench, for whenever the local-wizard build unblocks (still gated on his credential-flow call, Inbox item 31 — not re-asked).
Magnitude — open-source inference engine, YC S25
- Apache 2.0, launched ~2026-09-30. Instead of shipping precompiled kernels (llama.cpp style), it compiles and tunes them on the exact machine before the first token.
- Vendor numbers claim ~92% faster decode than llama.cpp on an M4 Pro Mac — vendor benchmarks, say so plainly; the HN thread has mixed practitioner reports plus real benchmark criticism. Treat as a claim, not a fact.
- Why it matters: a drop-in engine upgrade, not another model to evaluate. OpenAI-compatible API, one-click hookup to Hermes and others, fully offline after model downloads, macOS/Linux/Windows. The local-wizard build was specced on Ollama/MLX — worth Timmy benchmarking Magnitude against the spec when the build unblocks.
- Real caveats: two-person team, two months old, bug reports on record, no multi-GPU, small model catalog.
Liquid LFM2.5-24B
- ~24B total, ~2B active per token, 131k context, with MLX and llama.cpp support. Company-reported 184 tok/s on an M5 Max; stronger tool-use results, acknowledged weaker coding and knowledge.
- Worth Timmy's benchmark list given Alexander's Mac hardware — not blind adoption: his workload is code-heavy, and this one admits coding is its weak leg.
Context only (not a candidate)
- Meta's Muse Spark Open (816B total, 1M context) needs roughly two M5-Ultra Mac Studios — direction of travel for the open-weight race, nowhere near his hardware.
Open questions for Alexander via Timmy (Telegram, his call): none new — item 31 (credential flow) stands unasked. Magnitude's only action item is a benchmark comparison, which is Timmy's lane when the build resumes.