Open Model WeightsNVIDIA · FIELD VERIFIED
Primary sources last checked on 26 Oct 2026. Verification freshness is tracked separately from factual change, and unknown values remain explicit rather than inferred.
MODEL INTELLIGENCE · SOURCE-FIRST
Open Model Weights records Nemotron 3.5 Bolt E4B BF16 as a 7.99616B-parameter text assistant and tool use open-weight model in the Nemotron 3.5 Bolt family, with NVIDIA listed as the developer. The checked source repository provides 3 recognized weight artifacts, format evidence for Safetensors and a verified context window of 65,536 tokens. Values that are not evidenced in the checked sources remain explicitly undisclosed or unknown.
The repository declares NVIDIA OpenMDW License Agreement v1.1. NVIDIA commercial/non-commercial permission remains subject to OpenMDW-1.1 terms. Open Model Weights reports the checked license evidence and classification; this is not legal advice.
The checked sources declare google/gemma-4-E4B-it as lineage evidence. Training recipe status: training-described-in-model-card; data disclosure status: dataset-metadata-declared.
+ Save to my shortlistMonitor changes →Local browser monitoring · no background email alerts
DIRECT ANSWER · FEATURED MODEL BRIEFING
Source-grounded editorial layer · reviewed 2026-10-26
Nemotron 3.5 Bolt E4B BF16 is a first-party open-weight small local-agent language model released for fast text assistants, instruction following and multi-turn tools. The publisher documents BF16 is the full-precision publisher release; a Q4_K_M 4GB version and the DFlash drafter are different deployable artifacts. This specific version must pass source-verified model-file, model-card, exact SHA and license checks before Open Model Weights publishes a field-verified profile. Its declared 8B total / 4B effective and 64K text tokens are differentiated from measured tensors, runtime memory and independent performance. Vendor benchmark claims do not confer an automatic Leading designation. Open Model Weights independently checked the source revision, files and declared license for this canonical record; it has not executed model inference.
WHY IT MATTERS
BF16 is the full-precision publisher release; a Q4_K_M 4GB version and the DFlash drafter are different deployable artifacts.
NVIDIA identifies Gemma-4-E4B-it as its starting checkpoint and describes further RL post-training for structured outputs and tools.
NVIDIA reports 8B total and 4B effective model parameters, describing distinct storage and activated compute concepts.
MODEL-SELECTION VIEW
BF16 is the full-precision publisher release; a Q4_K_M 4GB version and the DFlash drafter are different deployable artifacts. The source-of-truth is the original publisher checkpoint and its revision-pinned file manifest. Keep the technical observation separate from release marketing, hosted-service features, third-party comparisons and any later edits to the model card; this is a prerequisite for accurate open-weight AI discovery.
Never size the full BF16 weight set using the 4GB label or treat an optional draft checkpoint as part of generic target inference. A reproducible evaluation must retain the original runtime version, model and dependent component revisions, task prompts, negative examples, latency and actual peak memory. This page documents a practical verification protocol, not an Open Model Weights independent inference result or universal model ranking.
NVIDIA identifies Gemma-4-E4B-it as its starting checkpoint and describes further RL post-training for structured outputs and tools. The source-of-truth is the original publisher checkpoint and its revision-pinned file manifest. Keep the technical observation separate from release marketing, hosted-service features, third-party comparisons and any later edits to the model card; this is a prerequisite for accurate open-weight AI discovery.
Assess real improvements with matched chat template, tokenizer, permitted tools and decoding parameters against the original foundation. A reproducible evaluation must retain the original runtime version, model and dependent component revisions, task prompts, negative examples, latency and actual peak memory. This page documents a practical verification protocol, not an Open Model Weights independent inference result or universal model ranking.
NVIDIA reports 8B total and 4B effective model parameters, describing distinct storage and activated compute concepts. The source-of-truth is the original publisher checkpoint and its revision-pinned file manifest. Keep the technical observation separate from release marketing, hosted-service features, third-party comparisons and any later edits to the model card; this is a prerequisite for accurate open-weight AI discovery.
Effective compute per turn does not imply the corresponding weights are the only tensors resident in device memory. A reproducible evaluation must retain the original runtime version, model and dependent component revisions, task prompts, negative examples, latency and actual peak memory. This page documents a practical verification protocol, not an Open Model Weights independent inference result or universal model ranking.
The model is intended for text interfaces and publisher serving examples disable input images, video and audio explicitly. The source-of-truth is the original publisher checkpoint and its revision-pinned file manifest. Keep the technical observation separate from release marketing, hosted-service features, third-party comparisons and any later edits to the model card; this is a prerequisite for accurate open-weight AI discovery.
A generic Transformers helper can display inherited multimodal examples; the checkpoint-specific serving contract remains authoritative. A reproducible evaluation must retain the original runtime version, model and dependent component revisions, task prompts, negative examples, latency and actual peak memory. This page documents a practical verification protocol, not an Open Model Weights independent inference result or universal model ranking.
NVIDIA publishes vLLM deployment settings, tool parsing configuration and constrained inference targets for local assistants. The source-of-truth is the original publisher checkpoint and its revision-pinned file manifest. Keep the technical observation separate from release marketing, hosted-service features, third-party comparisons and any later edits to the model card; this is a prerequisite for accurate open-weight AI discovery.
Measure tool-call errors, time to first token, peak device memory and full-context latency under pinned runtime dependencies. A reproducible evaluation must retain the original runtime version, model and dependent component revisions, task prompts, negative examples, latency and actual peak memory. This page documents a practical verification protocol, not an Open Model Weights independent inference result or universal model ranking.
NVIDIA describes permitted commercial and non-commercial use governed by OpenMDW License Agreement version 1.1. The source-of-truth is the original publisher checkpoint and its revision-pinned file manifest. Keep the technical observation separate from release marketing, hosted-service features, third-party comparisons and any later edits to the model card; this is a prerequisite for accurate open-weight AI discovery.
This is not Apache-2.0 or MIT; third-party Gemma ancestry and each redistributed conversion need their own notice review. A reproducible evaluation must retain the original runtime version, model and dependent component revisions, task prompts, negative examples, latency and actual peak memory. This page documents a practical verification protocol, not an Open Model Weights independent inference result or universal model ranking.
IMPLEMENTATION & EVALUATION
Four-step original-source test plan for Nemotron 3.5 Bolt E4B BF16, not an independently executed performance benchmark.
Record exact source SHA, original artifacts and OpenMDW-1.1 license text with upstream notices.
Immutable manifest with official revision and compatible dependency list
Upstream repository can change after intake
Run Nemotron 3.5 Bolt E4B BF16 through its original official task interface, retaining model and dependency versions.
Repeatable source-aligned output type and stable input handling
Generic family examples may not match this checkpoint
Score task-specific positive and negative cases for fast text assistants, instruction following and multi-turn tools, with published failure examples.
Held-out task quality table with traceable observed errors
Publisher benchmark numbers are not independent results
Record throughput, p95 latency, peak memory, operating system and rollback criteria under fixed tasks.
Written deployment decision with real observed resource metrics
Weight-only arithmetic omits cache, activations and extra decoder assets
Proposed tests, not OMW benchmark results. Read the official model card at the pinned source revision ↗ Inspect OMW field evidence ↗
PRIMARY & INDEPENDENT SOURCES
FAQ
Nemotron 3.5 Bolt E4B BF16 is a publisher-released small local-agent language model intended for fast text assistants, instruction following and multi-turn tools. Its official documentation describes BF16 is the full-precision publisher release; a Q4_K_M 4GB version and the DFlash drafter are different deployable artifacts. Open Model Weights verifies source records rather than inventing comparative scores.
The official publisher maintains a Hugging Face checkpoint with a model card and weight-file inventory. Its exact source SHA and artifact accessibility must pass a targeted verifier before a canonical record can be called field-verified.
The official repository declares OpenMDW-1.1, but the exact revision's license file, dependency scope and downstream obligations must be checked. The public-weight label does not grant blanket permission for any conceivable deployment.
Notable reflects source-released specialist significance, not an absolute AI leaderboard result. Independent checkpoint-equivalent benchmark results are not currently sufficient to award Leading status for this record.
There is no universally verified device-memory minimum for this deployment. Stored parameters, precision, caches, external decoders, tensor parallelism and inference frameworks all affect the actual peak resident footprint.
Pin a 40-character source SHA and matching model and component files, reproduce the official task interface and test against held-out positive and negative cases. Record real device memory, latency, quality failures and license conditions.
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
NVIDIA commercial/non-commercial permission remains subject to OpenMDW-1.1 terms
This section summarizes the checked repository evidence. It is not legal advice.
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.
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
02 · LICENSE
Classification describes the checked license terms; it is not legal advice.
03 · HARDWARE
Weight-only decimal-GB estimate from total parameter count; excludes KV cache, activations, runtime overhead, optimizer state and sharding.
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
Local assistant and tool model. A candidate for a reasoning or coordination role. No end-to-end agent performance is inferred from its model category.
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: NVIDIA OpenMDW License Agreement v1.1 (allowed-with-conditions commercial-use classification; see original terms).
Context: 65536 configured tokens.
BF16 weights: 15.99 estimated decimal GB, not runtime VRAM.
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 →