Open Model WeightsNVIDIA · FIELD VERIFIED
Primary sources last checked on 1 Nov 2026. Verification freshness is tracked separately from factual change, and unknown values remain explicit rather than inferred.
CONCEPTUAL MODEL BRIEF · ORIGINAL ILLUSTRATION
A topic-specific editorial visual, not an independent model performance measurement.MODEL INTELLIGENCE · SOURCE-FIRST
Open Model Weights records DeepSeek V4.1 Flash 1017 NVFP4 as a Mixed-precision NVIDIA quantized derivative; do not infer a new base-model parameter total from packed weight tensors-parameter text and images → text open-weight model in the DeepSeek V4.1 Flash family, with NVIDIA listed as the developer. The checked source repository provides 49 recognized weight artifacts, format evidence for Safetensors and a verified context window of 1,048,576 tokens. Values that are not evidenced in the checked sources remain explicitly undisclosed or unknown.
The repository declares mit. Commercial use allowed under mit conditions. Open Model Weights reports the checked license evidence and classification; this is not legal advice.
The checked sources declare deepseek-ai/DeepSeek-V4.1-Flash-1017 as lineage evidence. Training recipe status: training-described-in-model-card; data disclosure status: described-in-model-card.
+ Save to my shortlistMonitor changes →Local browser monitoring · no background email alerts
DIRECT ANSWER · FEATURED MODEL BRIEFING
Source-grounded editorial layer · reviewed 2026-11-01
NVIDIA's DeepSeek-V4.1-Flash-1017-NVFP4 is a derivative quantized checkpoint of the existing DeepSeek V4.1 Flash model, optimized for Blackwell-class inference. The model card declares MIT and a special patched vLLM path, not independent speed or quality results. Open Model Weights independently inspected the exact repository revision, listed weight artifacts and declared license; it has not run the model or reproduced publisher benchmark results. The repository includes a separately checked license file.
WHY IT MATTERS
NVIDIA publishes an optimized quantization of the existing DeepSeek-V4.1-Flash-1017 base checkpoint rather than claiming to train a new foundation model. It targets 384 routed experts across 40 transformer layers with W4A4 NVFP4 conversion.
NVIDIA says it preserved dequantized values for the converted expert blocks, apart from signed-zero canonicalization. It separately identifies the Engram conversion as lossy and describes fallback activation scale choices.
The model card names Blackwell GB300-class hardware and a patched vLLM implementation for Engram support. Standard out-of-the-box inference engines should not be assumed compatible until exact dependency versions and patches are verified.
MODEL-SELECTION VIEW
NVIDIA publishes an optimized quantization of the existing DeepSeek-V4.1-Flash-1017 base checkpoint rather than claiming to train a new foundation model. It targets 384 routed experts across 40 transformer layers with W4A4 NVFP4 conversion.
The original DeepSeek model should retain its canonical record and any independent evaluation. A quantized derivative needs its own exact artifact identifier so runtime choices can be tracked without double-counting the base model as a separate architecture.
The checkpoint occupies approximately 400 GiB in 48 Safetensors shards according to the publisher, compared with approximately 475 GiB for the cited source distribution. That is an artifact-specific storage statement, not a promise about inference VRAM.
Some components remain at their source precision. The Engram lookup tables undergo MXFP4 conversion with loss, while other source tensors use different dtypes; describing the entire model as uniformly four-bit would misstate the actual representation.
NVIDIA says it preserved dequantized values for the converted expert blocks, apart from signed-zero canonicalization. It separately identifies the Engram conversion as lossy and describes fallback activation scale choices.
Preserving converted weight values does not guarantee identical generated answers or deterministic end-to-end behavior. Compare base and variant checkpoints on matched multimodal tasks, sampling settings and outputs before making accuracy-equivalence claims.
The model card names Blackwell GB300-class hardware and a patched vLLM implementation for Engram support. Standard out-of-the-box inference engines should not be assumed compatible until exact dependency versions and patches are verified.
Latency, throughput, peak memory and quality will depend on hardware topology, cache sizes, batching, quantization kernels and attention implementation. Publisher numbers must remain attributed to their specific test setup.
The record matters for deployment engineering because it provides an official, explicitly named quantized derivative with reproducible artifacts. It should be cross-linked to the existing DeepSeek V4.1 Flash base model, never promoted as a distinct frontier foundation architecture.
Leading classification would require independent evaluation of this exact NVFP4 distribution, including accuracy trade-offs and hardware-specific performance. Source verification and publisher-only benchmarks are not enough.
Pin the NVIDIA repository revision, the original DeepSeek base revision, the vLLM patch set and the Blackwell hardware/software stack. Retain exact file manifests and kernel settings in any deployment report.
Test representative coding, visual, reasoning and agent tasks against the base distribution, then measure peak memory, latency, throughput, accuracy differences and failures on the actual intended hardware. Record when the special Engram path is required.
IMPLEMENTATION & EVALUATION
Reproducible operator checklist; these steps are not claims of an OMW model run.
A named model does not guarantee identical checkpoint revisions.
Record 40-character SHA, original file names, configuration and tokenizer/processor versions.
Unpinned revisions can silently change weights or usage requirements.
A specialist checkpoint can fail outside the publisher's demonstrated conditions.
Document actual prompt templates, modality preprocessing, inference backend and compatible software versions.
Using unsupported serving backends can invalidate comparisons.
Model card labels and code licenses are not automatically permission for every deployed use.
Check exact weight terms, notices, tool permissions, data privacy and runtime isolation.
Restricted weight rights or unsafe action tools can make a deployment unsuitable.
Vendor benchmark figures are useful evidence, but are not OMW-run observations.
Pre-register held-out tasks and compare baseline checkpoint, hardware, quality, memory and error rates.
Attributing first-party results to independent inference would mislead readers.
Proposed tests, not OMW benchmark results. Read the official model card at the pinned source revision ↗ Inspect OMW field evidence ↗
PRIMARY & INDEPENDENT SOURCES
FAQ
No. It is an NVIDIA-produced quantized derivative of an existing DeepSeek V4.1 Flash checkpoint, represented separately for reproducible deployments.
No. Only specified expert components are converted to NVFP4 W4A4; Engram tables use MXFP4 and other attention or vision components can retain other source formats.
NVIDIA describes roughly 400 GiB across 48 Safetensors shards, an artifact-specific count that should not be translated directly into peak serving VRAM.
Do not assume so. The first-party card describes a patched vLLM Engram implementation and specific Blackwell serving conditions.
No guarantee follows from weight-block conversion. Engram conversion is lossy, and model outputs depend on the complete runtime configuration.
No. Hardware and quality statements remain NVIDIA-attributed until exact independent measurements on the public checkpoint are available.
03 · PRACTICAL TECHNICAL PROFILE
Inspect weight-memory estimates, model lineage and runtime evidence separately from promotional or untested performance claims.
DEPLOYMENT PROFILE
These bars visualize weight-only decimal-GB estimates. They exclude KV cache, activations, runtime overhead, optimizer state and sharding.
LICENSE & USE
Commercial use allowed under mit conditions
This section summarizes the checked repository evidence. It is not legal advice. Inspect license evidence ↗
LINEAGE & PROVENANCE
Open Model Weights does not infer a hidden base model or training corpus when the checked sources do not declare one.
RUNTIME SIGNALS
Declared and model-card-mentioned runtimes are discovery signals. They are not treated as Open Model Weights runtime tests unless explicitly measured.
EVIDENCE TRAIL
This is where the article remains auditable: source links on one side, the registry's observed change history on the other.
Checked Hugging Face API metadata, exact repository file list, model card and config when accessible.
04 · EVIDENCE AND UPDATES
Inspect dated source checks and any differences between published claims and the evidence; these checks are not independent model benchmarks.
SOURCE-BASED VISUAL BRIEF · NOTABLE
Three evidence-linked visual summaries: parameter scale, illustrative precision memory and verifiable provenance. No performance ranking is implied.
No comparable single-model parameter total is available.
No reliable single-model weight-only memory figure.
FRESHNESS & VERIFICATION
A verification check does not pretend the model changed. Open Model Weights updates the verification date after a real primary-source check, while factual-update dates move only when retained source evidence changes.
CONFLICT INTELLIGENCE
05 · SOURCE RECORD
Full repository-level attributes, licensing, format details and provenance are available below without hiding or rewriting the original data.
01 · WEIGHTS
model-00001-of-00048.safetensorsmodel-00002-of-00048.safetensorsmodel-00003-of-00048.safetensorsmodel-00004-of-00048.safetensorsmodel-00005-of-00048.safetensorsmodel-00006-of-00048.safetensorsmodel-00007-of-00048.safetensorsmodel-00008-of-00048.safetensorsmodel-00009-of-00048.safetensorsmodel-00010-of-00048.safetensorsmodel-00011-of-00048.safetensorsmodel-00012-of-00048.safetensorsmodel-00013-of-00048.safetensorsmodel-00014-of-00048.safetensorsmodel-00015-of-00048.safetensorsmodel-00016-of-00048.safetensorsmodel-00017-of-00048.safetensorsmodel-00018-of-00048.safetensorsmodel-00019-of-00048.safetensorsmodel-00020-of-00048.safetensorsmodel-00021-of-00048.safetensorsmodel-00022-of-00048.safetensorsmodel-00023-of-00048.safetensorsmodel-00024-of-00048.safetensorsmodel-00025-of-00048.safetensorsmodel-00026-of-00048.safetensorsmodel-00027-of-00048.safetensorsmodel-00028-of-00048.safetensorsmodel-00029-of-00048.safetensorsmodel-00030-of-00048.safetensorsmodel-00031-of-00048.safetensorsmodel-00032-of-00048.safetensorsmodel-00033-of-00048.safetensorsmodel-00034-of-00048.safetensorsmodel-00035-of-00048.safetensorsmodel-00036-of-00048.safetensorsmodel-00037-of-00048.safetensorsmodel-00038-of-00048.safetensorsmodel-00039-of-00048.safetensorsmodel-00040-of-00048.safetensorsmodel-00041-of-00048.safetensorsmodel-00042-of-00048.safetensorsmodel-00043-of-00048.safetensorsmodel-00044-of-00048.safetensorsmodel-00045-of-00048.safetensorsmodel-00046-of-00048.safetensorsmodel-00047-of-00048.safetensorsmodel-00048-of-00048.safetensorsmodel.safetensors.index.json02 · LICENSE
Classification describes the checked license terms; it is not legal advice.
License / model-card evidence ↗03 · HARDWARE
Publisher describes a mixed NVFP4/MXFP4 derivative of roughly 400 GiB with a custom Engram-aware patched vLLM requirement. No OMW Blackwell VRAM/latency measurement.
04 · FORMATS
“Not observed” means no matching repository artifact/name/dtype signal was found; it does not prove no third-party conversion exists.
05 · LINEAGE
06 · TRAINING ASSETS
07 · RUNTIME SUPPORT
Open Model Weights separates model-repository evidence from external runtime-release evidence. An external support statement does not change the official model repository's weight formats and does not imply an Open Model Weights inference benchmark.
08 · SOURCES
09 · VERIFICATION HISTORY
Checked Hugging Face API metadata, exact repository file list, model card and config when accessible.
MODELS FOR AI AGENTS · ROLE EVIDENCE
Specialized model. A supporting component. This checkpoint does not, by category alone, serve as the planning and tool-execution controller.
The curated notable tier is not an agent-readiness rating. Model-specific tool-call success, long-horizon task accuracy and compatible runtime configurations are not independently established here.
Source-linked license: mit (allowed commercial-use classification; see original terms).
Context: 1048576 configured tokens.
BF16 weights: Not calculated.
Compare agent model candidates → · Agent evaluation method → · Inspect original model repository ↗
These records are separately sourced. They do not replace independent model-plus-runtime tests.
No model-specific agent capability evaluation is recorded for this checkpoint. Missing evidence is unknown, not a failed task.
Read evidence details and benchmarks → · Use the agent model selector →