OtherU publishes quantized DeepSeek-V4-Flash artifacts for Ember on AMD Strix Halo. Quantization reduces the memory required for model weights, allowing a large model to fit the reference unified-memory platform.
This page describes the earlier text-only DeepSeek-V4-Flash-Strix-Halo-GGUF artifact and its affine format. The default Ember deployment since 2026.9.3 uses the separate vision model repository. Use the artifact revision selected by your Ember release.
Processing stages
| Stage | Purpose |
|---|---|
| Acquire | Record the upstream model and input revisions |
| Transform | Apply any documented model edits before conversion |
| Convert | Produce a compatible GGUF representation |
| Quantize | Apply the per-tensor precision recipe |
| Validate | Check artifact integrity, tensor layout, and numerical behavior |
| Package | Publish compatible artifacts with their provenance and checksums |
These stages describe the artifact workflow. They are not a public one-command toolchain or a claim that every internal pipeline script has completed a full-model run.
Affine fp2
The text-only recipe uses the custom Q2_0_ROCMFP2 tensor format for routed experts. Its affine reconstruction is value = code × scale − offset. Storage overhead raises the effective precision above two bits per weight; the published recipe is described as 2.58 bpw.
Q2_0_ROCMFP2 and the similarly named Q2_0_ROCMFPX are not interchangeable. Their numerical encodings differ. A runtime must implement the tensor types in the artifact; the .gguf extension alone does not establish compatibility.
Artifact compatibility
Ember's release configuration identifies matching target-model, drafter, and, for vision, projector artifacts. Use the published checksums and immutable revisions to identify the files you deploy. Avoid combining components solely because their filenames look related.
Quality and measurements
Quantization and optional refusal-direction edits can change model behavior. A successful tensor-layout check or throughput benchmark does not establish preserved capability. Evaluate numerical error, task quality, and the behavior needed by your application separately.
The earlier artifact documentation did not establish a full capability assessment for its quantization recipe. Consult the model card and Ember's quality reports for evidence relevant to a particular revision. Performance results belong to their measured model, release, and workload.