license: other license_name: qwen-community-license-1.0 license_link: https://huggingface.co/Qwen/Qwen3.8-Flash-Next/blob/main/LICENSE language:
- en
- zh base_model:
- orcarouter/Qwen3.8-Flash-Next-Uncensored
- Qwen/Qwen3.8-Flash-Next base_model_relation: quantized pipeline_tag: image-text-to-text library_name: transformers tags:
- qwen
- qwen3.8
- Strata
- flash-next
- uncensored
- modelopt
- nvfp4
- gguf
Update README.md
Model for sergqwer/strata-nvfp4
-git (https://github.com/sergqwer/strata-nvfp4)
Installation
-Download and copy the contents to the models folder.
////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////
OrcaRouter Qwen3.8 Flash-Next — ModelOpt NVFP4 (sergqwer/strata-nvfp4)
Published by request from the Pennyroyal community. This is a ModelOpt NVFP4 conversion of OrcaRouter’s uncensored Qwen3.8 Flash-Next checkpoint—the conversion I use for local agentic work with Pennyroyal/SGLang.
I’ve also heard reports from several community members of it outperforming the stock model on agentic tasks, which added to the interest in sharing this working format.
The original model modification is OrcaRouter’s work. My contribution here is the quantization, compatible packaging and runtime testing—not additional fine-tuning or uncensoring.
Model lineage
- Modified source: OrcaRouter/Qwen3.8-Flash-Next-Uncensored
- Base checkpoint: Qwen/Qwen3.8-Flash-Next
- Runtime: Pennyroyal/SGLang
Quantization and packaging
The checkpoint uses NVIDIA ModelOpt 0.46.0 NVFP4: 4-bit floating-point weights and activations, group size 16, on selected Linear modules. Attention and Gated-DeltaNet paths, routers, shared experts, hyper-connections, PLE, vision, embeddings, LM head and MTP are excluded from NVFP4 according to the packaged configuration. The large PLE embedding shards are FP8 E4M3. It is stored as sharded safetensors and is approximately 135.3 GB (126 GiB). Pennyroyal's optional online-FP8 optimizations are runtime settings, separate from this on-disk conversion. RAM versus NVMe PLE placement is also a runtime choice; see NVMe PLE setup.
This is the working checkpoint used for my daily agentic workload, including thinking and tool use, and tested with Pennyroyal’s Flash-Next native-MTP/FR-Spec path on an RTX PRO 6000 Blackwell. That is practical runtime evidence, not a claim of new benchmark scores or broad hardware compatibility.
Run with Pennyroyal
For a native installation, download the repository to a local directory and set that directory as TARGET_MODEL; then follow Launch Flash-Next with FR-Spec.
For the container, mount the parent model directory as HOST_MODELS_ROOT and set TARGET_MODEL to its corresponding path below /models, for example /models/OrcaRouter-Qwen3.8-Flash-Next-Uncensored-ModelOpt-NVFP4. See Choose a profile and configure .env.
The supplied source files have conflicting license metadata: the source README labels the model Apache-2.0, while its included LICENSE is Qwen Community License 1.0. This upload preserves that license file and declares the Hub license as other; review the included terms before use.