# Qwen3.8-Flash-Next: Self-Hosting Hardware, Scenarios and Commercial License

> Cost-sensitive coding, office and multimodal agent workloads that need strong public benchmark results but can accept preview-stage serving software or a managed production derivative.

- Verified: 2026-09-12
- Released: 2026-08-26
- Canonical: https://chinaapi.ai/open-model-deployment/qwen3.8-flash-next/
- Back to directory: https://chinaapi.ai/open-model-deployment/
- Machine-readable dataset: https://chinaapi.ai/data/open-model-deployment.json

## Model facts

- Vendor: Alibaba Qwen
- Parameters: 125B main + 51B N-gram embedding + 4B MTP total; 6B per token active
- Native precision: Official BF16 checkpoint (335.28 GiB) and official FP8 checkpoint (172.78 GiB); community and NVIDIA NVFP4 conversions are separate artifacts
- Context: 256K tokens
- Category: General and agentic foundation models
- Modalities: text, image, video → text
- Serving paths: Transformers, llama.cpp, MLX, vLLM, SGLang, TokenSpeed
- Official weights: Official BF16: https://huggingface.co/Qwen/Qwen3.8-Flash-Next?utm_source=chinaapi&utm_medium=research&utm_campaign=open-model-deployment; Official FP8: https://huggingface.co/Qwen/Qwen3.8-Flash-Next-FP8?utm_source=chinaapi&utm_medium=research&utm_campaign=open-model-deployment
- Model documentation: https://github.com/QwenLM/Qwen3.8-Flash-Next?utm_source=chinaapi&utm_medium=research&utm_campaign=open-model-deployment
- Deployment and cost analysis: https://chinaapi.ai/insights/qwen3-8-flash-next-local-deployment-value/

## Deployment evidence

- **Official launch / minimum — evidence B:** vLLM validates the official 172.78 GiB FP8 checkpoint on 2x GB300 as the minimum supported FP8 topology. The 51B N-gram table also needs at least 51GB of host memory plus runtime headroom. This is the official-framework minimum, not a promise about useful concurrency at 262K context. Source: https://recipes.vllm.ai/Qwen/Qwen3.8-Flash-Next?utm_source=chinaapi&utm_medium=research&utm_campaign=open-model-deployment
- **200-person team — evidence B:** For the defined 200-seat profile, start load testing with the validated 4x H100 80GB FP8 worker and CPU-offloaded N-gram table. vLLM reports about 1,430 output tokens/s at concurrency 64 on a random 1,024-input/256-output workload, but a team service still needs a second failure-domain worker if availability matters and separate tests for real multimodal and long-context traffic. Source: https://recipes.vllm.ai/Qwen/Qwen3.8-Flash-Next?utm_source=chinaapi&utm_medium=research&utm_campaign=open-model-deployment
- **Commercial API — evidence B:** Use at least two independently deployable workers; vLLM recommends 4x GB300 FP8 per full-tray worker and also validates 8x H200 TEP8, 4x H100 with N-gram CPU offload, and 4x MI355X. Final topology depends on context, image/video mix and SLOs, and commercial MaaS requires a separate Qwen license before launch. Source: https://recipes.vllm.ai/Qwen/Qwen3.8-Flash-Next?utm_source=chinaapi&utm_medium=research&utm_campaign=open-model-deployment

## Scenario evidence

- **Coding and software-engineering agents — Public evidence:** Qwen reports 58.7 on DeepSWE 1.1, 62.5 on SWE-bench Pro and 81.0 on SWE-bench Multilingual. These are vendor-published evaluations, not ChinaAPI reproductions. Source: https://huggingface.co/Qwen/Qwen3.8-Flash-Next?utm_source=chinaapi&utm_medium=research&utm_campaign=open-model-deployment
- **Long-horizon office and tool work — Public evidence:** The official card reports 73.9 on CoWorkBench, 55.7 on JobBench and 73.5 on Toolathlon Verified, with the vendor positioning the model for coding and office tasks. Source: https://huggingface.co/Qwen/Qwen3.8-Flash-Next?utm_source=chinaapi&utm_medium=research&utm_campaign=open-model-deployment
- **Single-workstation research pilot — Third-party reproduced:** A third-party NVFP4 recipe reported a median 43.81 output tokens/s on one 128GB DGX Spark at 8,192-token context and single concurrency. It used a 101 GiB non-Qwen-official four-bit checkpoint and custom serving work, so it is evidence of feasibility, not an official minimum or production-capacity result. Source: https://forums.developer.nvidia.com/t/fitting-qwen3-8-flash-next-180b-onto-one-dgx-spark-44-tok-s-at-four-bits/382117?utm_source=chinaapi&utm_medium=research&utm_campaign=open-model-deployment

## Commercial-use check

- License: Qwen Community License 1.0 — https://huggingface.co/Qwen/Qwen3.8-Flash-Next/blob/main/LICENSE?utm_source=chinaapi&utm_medium=research&utm_campaign=open-model-deployment
- MaaS / hosted service: A separate Qwen license is required before commercial use by a licensee or affiliate that conducts a Model-as-a-Service or AI Work Assistant business. This condition has no revenue threshold in the published license.
- Attribution: Include the copyright and permission notice. A commercial product or service above 100M monthly active users or USD 20M monthly revenue must prominently display the model name in its user interface.
- Exceptions: Internal use is exempt from the separate-license condition only when the software, outputs and model capabilities are not made available to a third party. Mere relay to a third-party hosted model is excluded from MaaS; single-purpose tools, non-coding/non-office domain assistants, and assistants that are only a feature of another primary product are excluded from the defined AI Work Assistant category.

## Avoid or validate first

- Calling the 6B active-parameter count a 6B memory footprint
- Treating a patched community NVFP4 single-workstation run as the official deployment baseline
- Launching MaaS or an independent coding or office assistant before obtaining the separate commercial license

## Known limitations and open questions

- No ChinaAPI reproduction of the downloadable checkpoint and no like-for-like quality comparison with the hosted qwen3.8-flash production model
- The official vLLM recipe requires a dedicated day-zero image; SGLang documentation said support was not yet in a tagged release when reviewed
- The validated 4x H100 throughput uses random 1,024-input/256-output requests and does not establish coding-agent, vision, video or 262K-context latency
- One-million-token context requires static YaRN and may affect shorter-context quality

## Evidence boundary

Loading weights, completing a first forward pass and meeting a production latency/SLA are separate thresholds. The 200-person and commercial tiers require workload-specific measurement. License summaries are product research, not legal advice.

Generated by GENERATED BY scripts/gen_open_model_deployment.py.