Beam 501B quantized to MXFP4 runs full 1M context on 2× AMD MI300X at 89.4% GPQA
ben_burtenshaw · x · 2026-10-29
Reflection AI's Beam 501B shipped in NVIDIA-first formats, but a community quantization to MXFP4 — plus a fix to vLLM's ROCm path — now runs the full model with full 1M-token context on just 2× AMD MI300X.
- MXFP4 experts shrink from 1,003 GB to 278 GB (official FP8 is 506 GB, needing 4× MI300X).
- GPQA-Diamond holds at 89.4% vs 89.3% for official NVFP4.
- 1,462 output tok/s on 2 GPUs at 128 requests; 21× faster decoding at 120k tokens.
- Open weights, scripts, and evals on Hugging Face.
More from Infra
- Strix Halo-specific kernels speed up vLLM prefill by up to 31x at long context — asfbrz96 · 2026-10-29
- FuriosaAI quadruples valuation to $2.9B in new funding round to challenge Nvidia — dinabass · 2026-10-29
- Dev questions Anthropic ops: Claude Code source leaked to prod, backbone outages 'routine' — cuenca · 2026-10-29
- System design for real-time AI: the 10 core components, from Kafka to edge inference — goyalshaliniuk · 2026-10-29
- Modal Hits $1B Annualized Run Rate and Raises Series D Funding — charles_irl · 2026-10-29
- GCP grows 103% YoY to $30.7B, set to pass Azure as #2 cloud — SumitGup · 2026-10-29