← Back to all builds
Qwen3.8-Flash-Next
Alibaba's multimodal MoE and a preview of Qwen4. 125B plus 51B n-gram embeddings, 6B active, 262K native context. Open-weight.
Related builds
hermes-qwen-vllm
vLLM plugin that fixes compaction and thinking-mode bugs when running Qwen 3.8 27B as a local Hermes agent.
Qwen3.8-27B on RTX 3090
Inference stack for Qwen3.8-27B on a single RTX 3090: 195k context, 82 tps single request, 672 tps peak.
Qwen3.8-27B NVFP4 on DGX Spark
Day-one Qwen3.8-27B deploy: NVFP4, 262K context, one DGX Spark, one command. Full reproducible stack.
Qwen3.8-27B FP8 on 2x 3090
27B on two consumer 3090s at ~75-85 tok/s with NVLink. Tool calling, reasoning, vision, MTP. Full deploy open-sourced.
