Find the artifact behind a model name.

Start with the current source collection

Choose a publisher model card, then a model-specific deployment recipe and an original performance report if one exists. For DeepSeek V4.1 Flash, the 1017 card and current recipe identify the new checkpoint; the September community measurements in this collection still identify earlier weights. A newer recipe does not update the model identity behind a historical result.

Start with provenance

Open the publisher model card and the pinned revision. Record the model family, architecture and required tokenizer/chat template. Follow a conversion back to its original weights. A converter's repository is evidence about that conversion, not the original publisher's endorsement. Keep original, conversion and runtime-specific identifiers in separate fields.

Review access and terms separately

Download access is not permission for every use. Review the declared license and any additional model terms for your intended application. An open-download flag is a historical access observation; a gated flag may require an agreement. Unknown license scope remains unresolved. The catalog preserves the old review state and never turns it into a commercial-use approval.

Select the runtime artifact

For llama.cpp, verify the architecture and GGUF support in your pinned build. For Ollama, select the exact locally installed model and capture its digest through the model-list endpoint. A linked GGUF variant is not proof that an Ollama tag exists or that an import works. The archive deliberately makes no runtime compatibility recommendation.

Capture immutable identity

Save the original revision, conversion revision if applicable, weight-file checksum, runtime version and runtime model digest. Use the downloadable selection record to keep missing values explicit. File size and artifact format belong to that exact artifact, not the model family. A later update needs a new record rather than silently rewriting your tested identity.

Compare metadata without a leaderboard

Compare source provenance, revision, format, terms and explicit limitations. A context claim or linked conversion does not prove usable inference on a machine. Read the original source for model-specific conditions, and do not turn metadata or platform popularity into a speed recommendation.

Prepare a safe trial and recovery

Once a machine and artifacts are separately authorized, begin with a small public prompt and a recorded memory budget. Keep the previous working artifact available. If a load fails, preserve the failure condition and revert the test configuration; do not relabel the entire model family unsupported. Release of any eventual performance result still requires the publication review process.

Look for defaults that differ from the title

The MiMo RL recipe currently defaults to MOPD, the Nemotron Lightning recipe named BF16 defaults to NVFP4, and DeepSeek Pro’s default is the dated 0813 artifact. Record the full resolved repository rather than the navigation label. Preserve variant-specific version requirements as well: the Pro default requires a newer engine than the model-wide minimum.

Sources & change log

1. Ollama list models API ↗Checked 2026-10-17 · Installed-model digest capture.
2. llama.cpp v0.7.0 build documentation ↗Checked 2026-10-17 · Pinned source tag, CPU, Metal and CUDA build paths; not tested locally.

17 Oct 2026 · Revision 2. Source-based task guide and downloadable resource. Source documentation inspected; environment reproduction not run. Moving documentation may change after review.