Primary use stated or inferred cautiously from official documentation.
N
NVIDIA / nvidia/nemotron-3.5-bolt-e4b
Nemotron 3.5 Bolt E4B
Compact NVIDIA model released on October 23, 2026, post-trained from Gemma-4-E4B-it. It has approximately 8 billion total and 4 billion effective parameters; E4B does not mean an MoE with four billion active parameters. We group BF16 and GGUF Q4_K_M in one profile, with results separated by precision. The documented modality is text to text: targeting voice assistants does not make it an audio recognition or synthesis system.
01 / SUMMARY
The essential
Maximum input capacity when published by the source.
Documented maximum generation limit.
Verified availability channels.
Primary functional family and verified specialties.
Declared terms for API access, model weights, or self-hosting.
02 / TECHNICAL SHEET
Limits and integration
| API ID | nvidia/nemotron-3.5-bolt-e4b |
|---|---|
| Model type | Text and reasoning · Code |
| Access model | API and open weights |
| License | OpenMDW 1.1; Gemma 4 base under Apache 2.0 |
| Deployment | API or self-hosted deployment |
| Release | 23 OCT 2026 |
| Knowledge cutoff | Post-training: May 2026 |
| Entrance | Text |
| Exit | Text |
| Context window | 65,536 tokens shared between input and output |
| maximum output | The card states up to 64K; this is not 64K in addition to the context |
| Reasoning | enable_thinking and conversational_mode; configurable budget |
| Published tools | Features · Multi-turn tools |
| Structured Outings | Not published |
| Batch processing | Not published |
| Prompt cache | Not published |
| Fine-tuning | Not published |
| Verified platforms | NVIDIA Build · Hugging Face · vLLM · llama.cpp |
Conversations, interactive characters and device assistants with tool calling. The BF16 variant supports research, evaluation and adaptation with vLLM; the 4 GB distribution prioritizes local deployment with llama.cpp. Reasoning and conversational mode are controlled through the chat template. NVIDIA recommends temperature 1.0 and top-p 0.95. The host agent must validate arguments, permissions and tool responses before executing actions.
The BF16 configuration sets 65,536 positions and the GGUF recipe uses the same context. The card summarizes input and output as up to 64K each; it does not establish a combined 128K window. NVIDIA Build rounds to 66K. Inherited image and audio components in config.json and automatic Hugging Face examples do not prove validated multimodal capabilities for this release. The 4 GB designation does not guarantee total memory at maximum context: add runtime and KV cache.
03 / CAPABILITIES
What it can do
Orientation
Conversations, interactive characters and device assistants with tool calling. The BF16 variant supports research, evaluation and adaptation with vLLM; the 4 GB distribution prioritizes local deployment with llama.cpp. Reasoning and conversational mode are controlled through the chat template. NVIDIA recommends temperature 1.0 and top-p 0.95. The host agent must validate arguments, permissions and tool responses before executing actions.
Context
65,536 tokens shared between input and output · The card states up to 64K; this is not 64K in addition to the context
Tools and integration
Functions · Multi-turn tools
Access
NVIDIA Build · Hugging Face · BF16 / Q4_K_M
Input modalities
Text
04 / PRICES
Documented cost
| Concept | Worth | Unit/condition |
|---|---|---|
| Standard input | Free prototyping endpoint; commercial rate not published | Standard API rate |
| Cached input | Not published | Reading reused prefixes |
| Cache write or storage | Not published | The condition varies by provider |
| Standard output | Free prototyping endpoint; commercial rate not published | May include reasoning tokens |
| Batch input | Not published | Asynchronous processing |
| Batch output | Not published | Asynchronous processing |
NVIDIA Build offers a free prototyping endpoint with the identifier nvidia/nemotron-3.5-bolt-e4b. A commercial rate, universal quota or permanent free access has not been verified; numeric costs remain empty. Downloadable weights require your own infrastructure. The model uses OpenMDW 1.1 and the Gemma 4 base declares Apache 2.0. DFlash drafters are optional speculative-decoding components, not independent chat models or a guarantee of acceleration on every device.
i Prices change and may depend on level, region or context length. Check the source before making a decision.
05 / EVALUATIONS
How to read the results
“
NVIDIA publishes a NeMo Gym comparison separating BF16 and Q4_K_M. We record BFCL-v3, ToolSandbox and RULER at 64K: they evaluate function calling, stateful tools and context retrieval, respectively; they do not measure the same thing. Tests use reasoning except RoleBench, EQ-Bench v3 and RULER. These are not Inferama measurements or safety tests. Precision can change results in either direction; preserve checkpoint, template, sampling, hardware and evaluator version when reproducing them.
| Benchmark | Result | Metric | Source |
|---|---|---|---|
| BFCL-v3 | BF16: 68.51% · Q4_K_M: 67.31% | Higher is better | View source ↗ |
| ToolSandbox | BF16: 68.77% · Q4_K_M: 74.13% | Higher is better | View source ↗ |
| RULER (64K) | BF16: 86.99% · Q4_K_M: 85.76% | Higher is better | View source ↗ |
Inferama only highlights a “best result” when the metric, test set, configuration, and date allow for an equivalent comparison. The supplier's figures are presented as claims from its own source.
06 / VERSIONS
Chronology
Bolt E4B · BF16 / Q4_K_M
Published weights and available prototyping endpoint; variants grouped without duplicate profiles.
07 / SOURCES
Traceability
TECHNICAL SHEETBolt E4B Q4_K_M · pinned model cardhuggingface.co↗
TECHNICAL SHEETBolt E4B BF16 · pinned model cardhuggingface.co↗
CONFIGURATIONBF16 configuration and contexthuggingface.co↗
APINVIDIA Build · endpoint and identifierbuild.nvidia.com↗
LAUNCHNVIDIA announcement and availability dateblogs.nvidia.com↗
LICENSEOpenMDW 1.1openmdw.ai↗