drex-max open weights ship as a 27GB Q8_0 file, Mac M5 Pro output within 0.02 of H100 full model
testingcatalog · x · 2026-11-02
testingcatalog reports that the open-weights model drex-max v1 is on Hugging Face as a single 27GB Q80 file, runnable on Mac via Nace's llama.cpp and Ollama builds. On an M5 Pro with 48GB RAM, answer probabilities for a sample request were within 0.02 of the full model on an NVIDIA H100, and a single question can yield up to 255 candidate answers — a notable result for local inference fidelity.
More from Infra
- Google's TPU economics: Gemini 3.1 Pro costs 76% less per task than Claude Sonnet 5.5 — Beth_Kindig · 2026-11-02
- Users Say OpenAI and Anthropic's Usage Cuts Are Pushing Them to Local Models — SEND_ME_YOUR_ASSPICS · 2026-11-02
- Spot GPU prices to hit 3-5x contract rates, low-price compute sellers biggest losers — bindureddy · 2026-11-02
- Rumor: Musk to renegotiate TeraFab terms with TSMC, dropping self-build insistence — SumitGup · 2026-11-02
- Crusoe CEO says neocloud debt fears overblown amid 'tidal wave' of compute demand — nathanbenaich · 2026-11-02
- Kapitsa hits 346 tok/s on a 35B MoE and runs a 180B model on 48GB Macs via SSD offloading — DotaMate · 2026-11-02