Qwen3.8-Flash-Next Hits 65 Tokens/Second on 12GB VRAM

A developer built a custom inference engine, Strata, running Qwen3.8-Flash-Next on consumer hardware including a 12GB RTX 5070, achieving about 65 tokens/s output and 430 tokens/s prompt processing, a major gain over the earlier llama.cpp setup. The engine is open-sourced, one-click installable, and currently optimized for CUDA only.

• Qwengram-0.8B: n-gram memory transfer cuts perplexity 5.05%2026-09-25
• Ling 3.0 Tiny brings edge intelligence to old laptops2026-09-26
• Mica v0.1 4B Builds an Iron Pickaxe in Minecraft With Zero Generated Tokens2026-09-26
• Swift1.5-Qwen3.8-Flash-Next shines in benchmarks2026-09-25
• Benchmark: ThinkingCap vs Swift vs Qwen 3.8-27B2026-09-25