Strata - Qwen 3.8 Flash Next - boosting decoding speed Tip
I tested on 4 consecutive prompts(Slovak language) in chat.
Proof1 community thread
Key points
- I've got some nice boost by increasing --adapt-every value from calibrated 1 to 6.
- Prompt 1-3 are simple questions, 4. is describing source code(20K tokens)
- [prompt number / ~avg. output length]
- [2 / 565] 51.9 -> 64.3 t/s (+24%)
Sources (1)
- [1]Strata - Qwen 3.8 Flash Next - boosting decoding speed Tipr/LocalLLaMA (top, daily) · Oct 17, 10:31 AM
I tested on 4 consecutive prompts(Slovak language) in chat.
I've got some nice boost by increasing --adapt-every value from calibrated 1 to 6.
Extractive summary: sentences quoted from the sources.
Before this
- Oct 17, 2026deepseek-ai/DeepSeek-V4.1-Flash-1017
- Oct 17, 2026Nemotron 3.5 Bolt, Muse Spark Open, and a Fast LFM2.5 24B
- Oct 16, 2026I don't get the qwen 3.8 flash next hype
- Oct 15, 2026capy-flash: Qwen3.8-Flash-Next (125B MoE) in full NVFP4 on two 16 GB consumer GPUs
- Oct 15, 2026huggingface/transformers v5.20.0: Release v5.20.0
- Oct 15, 2026Prompt tips from Anthropic
