Running Qwen3.8-Next-Flash on 96GB RAM: offload the n-gram table to SSD

Iory1998 · reddit · 2026-08-30

A user with dual 5070 Ti + 3090 (40GB VRAM) and 96GB RAM shares a hands-on debugging story running Qwen3.8-Next-Flash Q4 locally:

• Unsloth quants failed: Q4KXL (105GB) and IQ4XS (90GB) bake the 56B n-gram table into the weight files, so everything loads into memory; with KV cache, conversations beyond 60K–100K tokens crash llama.cpp silently, at only 12–14 t/s.
• The AtomicChat build is the key difference: it shards the n-gram table onto SSD, needing only 54GB of fast memory. The table is touched sparsely (2.7KB per token via hash), so SSD offload costs almost nothing.
• Results: a real 217K-token conversation ran end-to-end without crashing, full 256K context fits, and speed rose to 22 t/s. Prefill remains hardware-bound at 18 min for 220K tokens.
• Fun fact: the AtomicChat option was suggested by DeepSeek running benchmarks on his own rig. The author cautions quant quality parity with unsloth isn't guaranteed.

Original post →

More from Infra

• Grok claims 816B MoE model runs locally on dual M5 Ultra Mac Studios via MLX — grok · 2026-10-16
• 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

Infra channel →