Qwen3.8-Flash-Next-FP8 is a release of FP8-quantized model artifacts that lets practitioners run the Qwen3.8-Flash-Next architecture with much lower memory and storage cost while keeping evaluation performance nearly unchanged. The practical value is that teams can experiment with the 125B multimodal MoE architecture and its 262K→1M long-context capabilities without the full-precision hardware footprint.
Key Capabilities
- FP8 fine-grained quantization (block size 128): reduces model size and memory bandwidth needs so inference and hosting costs drop substantially while maintaining near-original metrics.
- Hugging Face Transformers-compatible artifacts: works with mainstream inference stacks (Transformers, vLLM, SGLang, TokenSpeed), enabling straightforward integration into existing pipelines.
- Multimodal + ultra-long context design: targets a causal LM with vision encoder and native 262,144 token context (extensible to 1,000,000), making it suitable for long-horizon agents, document reasoning, and long-video understanding.
- Architecture-preview of Qwen4 family: exposes hybrid attention (Gated DeltaNet + Qwen Sparse Attention), gated residuals, n-gram embeddings and MoE design so researchers can evaluate architectural trade-offs.
Who it's for and trade-offs
Great fit if you need to prototype or deploy a large multimodal LLM with far reduced memory/storage footprint and want compatibility with popular inference engines. Also useful for benchmarking agentic, coding and long-horizon tasks with realistic model behavior. Look elsewhere if you require the absolute highest-fidelity, unquantized weights for research into tiny numerical differences, or if your infra cannot yet support FP8 runtimes or the recommended serving engines. The release lowers hardware requirements but still assumes significant compute for production-scale throughput.
