Qwen3.8 Flash Next R9V IQ4_XS

Ready-to-arrange model bundle for the R9V dual-RDNA4 inference profile.

Contents and provenance

• target/: Unsloth UD-IQ4_XS target GGUF shards from unsloth/Qwen3.8-Flash-Next-GGUF.
• vision/: ggml-org Q8_0 vision projector. This projector was not quantized by R9V.
• mtp/: R9V-assembled minimal MTP checkpoint. Dense/nonexpert tensors come from the official BF16 checkpoint; routed experts come from the official block-FP8 checkpoint. R9V did not train these weights.
• metadata/: official Qwen tokenizer, processor, and model configuration.
• manifests/: the reference dual-R9700 hot-expert placement.
• sources.lock.json: exact upstream revisions, sizes, and hashes.

Unsloth and ggml-org are credited for the target quantization and vision projector. Qwen remains the model author and upstream rights holder.

PLE table

The 26.82 GiB per_layer_token_embd.weight payload is already present inside target shard 2 and is intentionally not uploaded again. Extract it to the fast SSD with R9V's metadata-driven tool. The extraction utility is shipped in the R9V source repository, not this model repository; follow the complete build, extraction, and launch instructions.

Reference configuration

• Two Radeon R9700 32 GiB GPUs, TP2.
• 128 GiB DDR5.
• MTP depth 2.
• 131,072-token capacity, BF16 QSA KV.
• Q8 vision input, one image/request.
• SSD PLE and tiered expert placement.

The immutable 22-file model package is public and remotely hash-verified. The R9V runtime has passed its reference-machine performance qualification and remains a release candidate until the documented package installation passes from a clean host. See the R9V qualification report for exact benchmark cells and protocol.

License

These model artifacts are distributed under Qwen Community License 1.0; see LICENSE. The R9V Apache-2.0 code license does not apply to model weights. Users are responsible for reviewing the Qwen license, including its separate terms for certain commercial MaaS/AI-work-assistant uses and scale thresholds. Artifact attribution and exact upstream revisions are recorded in THIRD_PARTY_NOTICES.md.