Qwen 350K Context Tested on M5 Max: Performance and Quality
Artistic_Okra7288 · reddit · 2026-08-30
Benchmark of 2-bit quantized Qwen3.8-Flash-Next running via llama.cpp on a 128GB M5 Max, utilizing a 358K-token context slot. Results show that while prefix reuse keeps prefills fast, a cold prefill after idle took 333s. Decode throughput tapers from 30-35 t/s at small context to 11.5 t/s at 169K tokens. The model performed strongly up to 100K tokens but exhibited role confusion at extreme depths, likely due to 2-bit quantization or preview model limitations.
More from Infra
- Same 1,000 Cloudflare Durable Objects: $4,050/month cut to $22 with one trick — threepointone · 2026-10-16
- NVIDIA ships FP4-quantized Qwen3.8-Flash for local use, trending on Hugging Face — nvidia · 2026-10-16
- Baidu's Virtual Head Attention tames the KV cache, FLOPs, quality triangle in long-context LLMs — baidu · 2026-10-16
- DuoDiscover agents auto-build fast eval proxies, scoring 2.68x TPU kernel speedup — berkeley_ai · 2026-10-16
- Google Raises 2026 Capital Spending to Up to $205 Billion — emmanuelvivier · 2026-10-16
- Google open-sources AX, a Kubernetes-style orchestrator for billions of AI agent tasks — thisdudelikesAI · 2026-10-16