Qwen 3.8 Flash Next rumored: 125B + 51B N-gram lookup tables, launches tomorrow
teortaxesTex · x · 2026-08-25
Rumors say Qwen 3.8 Flash Next (125B params + 51B N-gram embeddings, 6B active, based on next-gen Qwen 4 architecture) launches tomorrow. superelesha analyzes the 51B N-gram tables: a massive trainable lookup-memory that hashes recent token N-grams and mixes fetched vectors into the hidden state, so the main model spends less effort on frequent local patterns.
Memory math (ideal 4bpw):
- 125B main weights ≈ 58.2 GiB
- 51B N-gram tables ≈ 23.7 GiB, 82 GiB total
- Leaves only 14 GiB spare on 4×3090s
Caveat: real Q4 is usually heavier, and some quantizers keep embedding tables in FP16/BF16 — the dtype of the 51B parameters is the key question of this release.
More from Infra
- SK Hynix ties new US plant to looser chip equipment exports to China, sources say — pstAsiatech · 2026-10-15
- Perplexity's new embedding models: index with a 9B model, search with 0.6B — antoine_chaffin · 2026-10-15
- Unsloth ships GGUF quants for Nemotron 3.5: Bolt-74B runs in 24GB RAM — danielhanchen · 2026-10-15
- vLLM-Omni ships Day-0 serving for FLUX 3 Image Klein with BF16, FP8, NVFP4 support — vllm_project · 2026-10-15
- NVIDIA's Nemotron 3.5 Super VL lands training and inference support on Applied Compute — ypatil125 · 2026-10-15
- CoreWeave bond yield tops 16% as loans outlast customer contracts by ~2 years — tengyanAI · 2026-10-15