Nvidia listed nvidia/DeepSeek-V4-Pro-0813-NVFP4 on Hugging Face, packaging DeepSeek V4 Pro weights in NVFP4 for lower-memory inference on supported GPUs. The model page appeared on Hacker News in November 2026.
This is not a repeat of August's V4 Pro launch. The Hugging Face card dates to August–September 2026; November HN attention is what brings it back for builders who missed the NVFP4 artifact.
What changed versus August 2026
| Layer | August V4 Pro launch | Nvidia NVFP4 HF repo |
|---|---|---|
| Announcement owner | DeepSeek product/API | Nvidia Hugging Face org |
| Artifact | API, pricing, features | Quantized weights |
| Goal | Frontier coding model access | Efficient self-host inference |
| Precision | Full precision serving | NVFP4 compressed weights |
Builders who only use DeepSeek's API can ignore this repo. Teams with Blackwell-class or other FP4-capable clusters should evaluate VRAM savings.
NVFP4 in one paragraph
NVFP4 is Nvidia's 4-bit FP format for tensor cores that implement FP4 paths. It trades numerical precision for capacity and throughput, similar in role to Qwen3 NVFP4 multimodal releases explainx.ai already covered.
Always run your eval suite after quantization — coding, tool-use, and long-context tasks degrade differently.
How to try the weights
- Open huggingface.co/nvidia/DeepSeek-V4-Pro-0813-NVFP4 and read hardware requirements.
- Match driver, CUDA, and inference server versions Nvidia documents for FP4.
- Compare latency and quality against GLM NVFP4 vLLM guides patterns if you use vLLM.
- Keep August API path as fallback when quantized self-host fails tasks.
Community harness context
explainx.ai's J-Space Cognition Suite post discussed community harnesses around V4 Pro. A quantized checkpoint may change harness scores — rerun before you trust leaderboard claims.
Export-control and licensing reminders
DeepSeek models carry US policy and enterprise procurement scrutiny covered in China distillation reporting. Weights on Hugging Face do not remove compliance review for your jurisdiction.
When to skip this repo
- No FP4 GPU in your fleet.
- You need exact-match API behavior from DeepSeek-hosted endpoints.
- Your task requires maximum precision (some finance or safety evals).
Hardware planning table (conceptual)
| Constraint | NVFP4 self-host | API |
|---|---|---|
| CapEx GPUs | High upfront | Low |
| Per-token opex | Power + staff | Vendor bill |
| Privacy | Strong if air-gapped | Policy dependent |
| Update cadence | You merge weights | Vendor ships |
Quantized model weights and GPU memory footprint for NVFP4 DeepSeek V4 Pro inference
Image: Library reference — quantized checkpoints demand disciplined load and eval scripts.
TensorRT and server stack notes
Nvidia typically pairs NVFP4 artifacts with TensorRT-LLM or vLLM plugins documented in TensorRT 11 guides. Before you pin production, run A/B against BF16 baselines on:
- Multi-file repo edits (V4 Pro's strength in August launch marketing)
- JSON tool calls with strict schemas
- 128k+ context if the card supports it
Quantization surprises often appear only on long-context needle tests.
Memory math (illustrative, not vendor quotes)
Exact VRAM depends on parallelism and KV cache dtype. Teams use NVFP4 when a single H100-class GPU must host a frontier coder that BF16 would OOM. Measure your batch size; do not extrapolate from blog tables.
Security and supply chain
Download only from nvidia/ org on Hugging Face. Verify commit hashes in internal SBOMs. Hugging Face security tradeoffs apply to every weight pull.
Relation to Nvidia open-model recipes
See Dynamo recipes for Korean open models for how Nvidia bundles reference inference stacks. This DeepSeek NVFP4 card is another artifact in that ecosystem play: make frontier-class names runnable on Nvidia silicon.
Related reading
- DeepSeek V4 Pro August launch
- J-Space V4 Pro harness
- Nvidia Qwen3 NVFP4 multimodal
- GLM NVFP4 vLLM guide
- Nvidia Qwen3 8 Flash local 96GiB
- Hugging Face open source security tradeoff
Benchmark expectations after quantization
MMLU, HumanEval, and SWE-bench scores from August marketing assumed full-precision serving. Expect small regressions on numeric-heavy coding tasks when you move to NVFP4. Document acceptance thresholds before you flip traffic.
Run regression suites nightly if you auto-pull Hugging Face revisions — Transformers integration contracts posts show how silent tokenizer fixes break pipelines.
Multi-GPU deployment sketch
Large V4 Pro-class models typically need tensor parallel across multiple GPUs even with FP4. Follow patterns from GLM NVFP4 two-DGX guide:
- Shard weights across NVLink nodes.
- Pin NVFP4 kernels in vLLM or TRT-LLM config.
- Warm up with representative batch sizes.
- Monitor KV cache growth on long agent loops.
Licensing and attribution
DeepSeek licenses and Nvidia redistribution terms live on the model card. Enterprise legal should confirm redistribution rights if you embed weights in appliances you sell.
When Nvidia publishes more NVFP4 families
This repo landed beside Qwen3 NVFP4 cards. Pattern suggests Nvidia will NVFP4-wrap every frontier name customers request on Blackwell — treat HN front-page as ops signal, not product launch.
Cost worked example (hypothetical)
Suppose BF16 needs 8 GPUs and NVFP4 needs 4 for the same throughput target on your harness. Savings are hardware rent, not zero ops — quantization engineering and eval staff still bill hours.
FAQ for platform engineers
Should we mirror weights internally? Yes, for air-gapped CI that cannot hit Hugging Face every deploy.
Does this replace Nemotron for coding? Only if your evals say so — Nemotron coding recipes may still win on latency.
Will DeepSeek API prices drop? Unrelated to Nvidia HF repo — watch DeepSeek pricing pages, not this artifact.
Historical context: August versus November artifacts
When V4 Pro launched, teams debated API versus self-host economics in terminal bench posts. November's NVFP4 card is the self-host half catching up — useful if API rate limits or data residency block you.
Observability hooks
Log per-request precision path (fp4 vs fallback) in your gateway. When quality tickets arrive, you can tell whether quantization or prompt drift caused regression.
Container and CI sizing
Multi-hundred-GB downloads affect CI cache budgets. Mirror to internal artifact registry and pin sha256 in Terraform modules so reproducible deploys do not re-pull on every pod restart.
Partner ecosystem
CoreWeave, Lambda, and Nebius may expose managed V4 Pro NVFP4 before you bare-metal it. Compare $/token after staff cost — AI VPS pricing claimed-vs-verified reminds that rented GPUs shift monthly.
Research angle
Quantization papers often report perplexity; production cares about tool success rate. Publish your internal tables if you run serious A/B — the HN crowd wants numbers, not ad copy.
Edge cases
- Fine-tunes atop NVFP4 base may be unsupported — check trainer docs.
- Multi-modal V4 variants may need different cards — read tags on Hugging Face.
- Export rules if you ship appliances outside US — legal, not ML.
Step-by-step first deploy (checklist)
- Provision a FP4-capable GPU VM with matching driver.
- Clone
nvidia/DeepSeek-V4-Pro-0813-NVFP4with git lfs. - Install vLLM or TRT-LLM version Nvidia documents for FP4.
- Run a smoke prompt from August V4 Pro examples.
- Compare output to API on same prompt.
- Enable Prometheus GPU metrics before any user traffic.
Failure modes we have seen elsewhere
OOM during weight load usually means wrong parallel sharding. Garbage outputs often mean wrong tokenizer revision. Slow tokens may mean FP4 kernels disabled and silent BF16 fallback.
Why Nvidia publishes DeepSeek-quantized cards
Chip demand follows runnable model names. NVFP4 artifacts make DeepSeek a reference workload on Blackwell marketing slides, similar to Qwen cards. That is distribution strategy, not endorsement of every downstream use.
Closing builder takeaway
Download the card when you already committed to self-host V4 Pro and need VRAM relief. Stay on API if eval staff is thin or compliance forbids weight custody. The HN post is a pointer, not a mandate.
Alignment with Nvidia Dynamo and NCCL ops
Teams already running Dynamo recipes should verify whether FP4 kernels integrate with existing NCCL groups or need separate serving pools. Mixing precision paths in one Kubernetes Deployment without node selectors causes nondeterministic OOMs.
Documentation debt to close internally
After first deploy, write an internal runbook: model revision, GPU SKU, tokens/s, quality delta vs API, and rollback command. Future you will need it when Hugging Face bumps metadata without a semver tag.
When to wait for the next quant release
If V4 Pro ships a 0814 checkpoint with bugfixes, this 0813 NVFP4 artifact may stagnate. Watch Nvidia org on Hugging Face for new revisions before you freeze production.
Related internal link for coding agents
If coding agents pull models from Hugging Face automatically, add allow-list policy for nvidia/* repos so interns do not mix NVFP4 and BF16 checkpoints in the same Compose file by mistake.
Schedule a quarterly re-eval because FP4 kernel quality improves with each driver release, and stale internal runbooks become wrong quickly on fast-moving Blackwell fleets. File a ticket if tokens per watt does not improve after quantization on your standard batch size and context length.
Keep a pinned driver version in staging until production validates FP4 quality.
Read the Hugging Face model card for license terms and exact GPU requirements before any production deploy.
