Qwen3.8-Flash-Next Mac optimization: Linear sparse attention, SSD streaming, and custom Q4
memeka · reddit · 2026-08-30
Developer achieved extreme optimization for Qwen3.8-Flash-Next on M1 Max 64GB, enabling SSD streaming for tensors, engrams, and MTP. Key techniques include:
- Custom Q4 Quant: Spliced tensors from multiple Unsloth/AtomicChat quants for best performance.
- Metal Sparse Attention: Custom mechanism achieving near-linear degradation vs standard quadratic.
- Dynamic MTP: Automatically disables MTP when context size makes it counter-productive.
- Tradeoffs: Enabling MTP uses more RAM, dropping prefill from 180 tps to 170 tps, but boosting decode by 70% to 22 btps.
Author credits Claude for providing a week's worth of tokens and credits.
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