NVIDIA releases Nemotron 3.5 Bolt E4B with 8B parameters, a 4GB local variant and agent-focused reinforcement learning
NVIDIA released Nemotron 3.5 Bolt E4B on October 23, 2026, introducing a compact language model for local assistants, multi-turn tool use and embedded interactive applications. Derived from Google Gemma 4 E4B-it and post-trained by NVIDIA, it has approximately eight billion total parameters and four billion effective parameters.
The release supports up to 64,000 tokens of context and ships in BF16 and Q4_K_M variants, including an approximately 4GB quantized checkpoint. NVIDIA reports improvements on instruction-following and agent-interaction evaluations, although competitors lead on several long-context tests.
The model is distributed through NVIDIA's official repositories under OpenMDW License Agreement version 1.1. The 4GB designation concerns quantized model weights and is not a universal total-system-memory requirement.
··········
RELEASE DATE, MODEL VERSIONS AND AVAILABILITY.
NVIDIA's official documentation dates Nemotron 3.5 Bolt E4B to October 23, 2026. Its BF16 variant serves customization, evaluation and further training, while the Q4_K_M release targets memory-efficient local inference.
Both checkpoints share the Gemma 4 E4B architecture and NVIDIA's additional post-training. The model is an adaptation of an existing foundation checkpoint, not a freshly pretrained architecture.
........
Property | BF16 release | 4GB release |
Release date | October 23, 2026 | October 23, 2026 |
Total / effective parameters | 8B / 4B | 8B / 4B |
Architecture | Gemma 4 E4B | Gemma 4 E4B |
Precision | BF16 | Q4_K_M |
Context | 64K | 64K |
License | OpenMDW 1.1 | OpenMDW 1.1 |
........
The BF16 checkpoint retains substantially more weight precision; deployment memory also depends on inference overhead, runtime configuration and context length.
··········
GEMMA 4 E4B ARCHITECTURE AND EFFECTIVE PARAMETERS.
The model contains approximately eight billion total parameters but is described as having four billion effective parameters. These figures are not interchangeable: its entire checkpoint contains more weights than a conventional claim of four billion total parameters would imply.
An eight-billion-parameter BF16 checkpoint requires roughly 16 billion bytes of raw weight storage before KV cache, activation memory and runtime buffers. This calculation is a theoretical weight estimate rather than a measured minimum VRAM requirement.
The quantized Q4_K_M release stores weights in a compact format, but total application memory increases with context length, batching and other local components. Bolt E4B is text-in and text-out, not a native image or video model.
........
Property | Reported value | Meaning |
Base checkpoint | Gemma-4-E4B-it | Further post-trained |
Total parameters | 8B | Complete inventory |
Effective parameters | 4B | Architecture-specific |
BF16 raw weights | ~16 GB | Excludes runtime |
Quantized artifact | ~4 GB | Not entire system memory |
Input / output | Text / text | No native vision claimed |
........
··········
REINFORCEMENT LEARNING FOR AGENT INTERACTION.
NVIDIA further post-trained Gemma 4 E4B-it using reinforcement-learning environments designed to improve instruction following, tool-use behavior, structured output and multi-turn dialogue.
The published model card identifies Group Relative Policy Optimization (GRPO) and the NVIDIA NeMo RL and NeMo Gym toolchain. Its asynchronous training approach separates response generation from optimization, useful when simulated tool calls have variable completion times.
Agent benchmarks evaluate behaviors beyond factual answering: correct tool selection, argument formatting, procedure adherence, persona consistency and the ability to incorporate tool results. NVIDIA reports a development period from December 2025 to May 2026 and a May 2026 post-training data cutoff.
NVIDIA does not publish a complete training-token count, compute budget or exhaustive provenance breakdown for the training data.
··········
CONFIGURABLE REASONING AND CONVERSATIONAL MODES.
The model exposes configuration for reasoning and conversation. Developers can use enable_thinking to activate or disable reasoning behavior and conversational_mode to select interactive dialogue behavior. A runtime thinking budget can control generated reasoning computation.
Greater reasoning computation can benefit complex tasks but also increases generated tokens and latency. Interactive applications may prioritize shorter response times over elaborate deliberation.
NVIDIA documents temperature 1.0 and top_p 0.95 as recommended inference settings, rather than universally optimal values.
........
Setting | Function | Trade-off |
enable_thinking | Reasoning on/off | Compute and latency |
conversational_mode | Dialogue behavior | Task suitability |
Thinking budget | Reasoning allocation | Speed versus depth |
Temperature | 1.0 recommended | Sampling randomness |
top_p | 0.95 recommended | Nucleus sampling |
........
··········
TOOL CALLING AND ASYNCHRONOUS WORKFLOWS.
Bolt E4B supports structured tool calling through an OpenAI-compatible chat format. The application must still execute calls, validate results and enforce permissions; model-generated tool arguments are not authorization.
NVIDIA describes an asynchronous task pattern with persistent task identifiers and pending statuses, so external long-running operations can finish before results are returned to the model.
A production agent needs timeouts, retries, approval controls, authentication and recovery from partially completed operations. Bolt E4B itself is not a complete distributed job scheduler.
··········
THE 4GB QUANTIZED MODEL AND LOCAL INFERENCE.
The Q4_K_M checkpoint reduces stored weight size through lower-bit quantization. It is packaged as a 4GB model for deployment on memory-constrained local systems and for applications sharing hardware with other workloads.
Quantization trades some numerical precision for smaller storage. Benchmark changes are task dependent rather than uniformly negative or positive. Inference also consumes KV-cache memory, activations and execution buffers, especially with longer prompts or simultaneous sessions.
A local voice assistant, for example, still requires memory for speech recognition and synthesis. Local execution may reduce dependency on hosted APIs, but actual cost depends on hardware, electricity, utilization and maintenance.
........
Consideration | BF16 | Q4_K_M |
Weight precision | 16-bit | Quantized |
Weight footprint | Higher | Lower |
Primary use | Customization | Local inference |
KV cache | Additional | Additional |
Quantization effect | None from Q4 | Task-dependent |
........
··········
OFFICIAL BENCHMARKS: CONVERSATION AND AGENTS.
NVIDIA provides BF16 evaluation results against Gemma 4 E4B-it and Qwen3.5 4B using a NeMo Gym-based internal harness. The following scores are reported by NVIDIA, not independently reproduced.
........
Benchmark | Gemma 4 E4B-it | Qwen3.5 4B | Bolt BF16 |
MultiChallenge | 42.08 | 51.76 | 53.10 |
RoleBench | 49.95 | 53.66 | 67.05 |
RoleMRC | 74.62 | 65.81 | 80.63 |
EQ-Bench v3 | 41.17 | 32.92 | 39.40 |
RULER 64K | 86.55 | 95.15 | 86.99 |
LongMemEval | 70.95 | 66.85 | 72.10 |
LongBench v2 48K | 49.34 | 59.93 | 50.50 |
SysBench | 65.85 | 67.34 | 67.88 |
GaRAGe | 34.73 | 24.66 | 42.61 |
RAGTruth | 76.97 | 75.60 | 77.59 |
RelyToolBench | 61.60 | 43.35 | 63.16 |
........
Bolt scores 67.05 on RoleBench compared with 49.95 for its Gemma base, and 53.10 on MultiChallenge versus 42.08 for Gemma. It also leads the listed RelyToolBench and GaRAGe comparisons.
Performance is not uniformly superior. Qwen3.5 4B scores 95.15 on RULER 64K, versus 86.99 for Bolt, and 59.93 on LongBench v2 48K, versus 50.50. Different model-card and NIM evaluation configurations must not be combined casually.
··········
HOW QUANTIZATION AFFECTS EVALUATION.
NVIDIA reports evaluations for its Q4_K_M checkpoint as well as BF16. Lower-precision weights retain much of the measured capability but cause benchmark-specific changes.
........
Benchmark | Bolt BF16 | Bolt Q4_K_M |
MultiChallenge | 53.10 | 55.60 |
RoleBench | 67.05 | 71.21 |
RoleMRC | 80.63 | 79.12 |
EQ-Bench v3 | 39.40 | 38.15 |
RULER 64K | 86.99 | 85.76 |
LongMemEval | 72.10 | 69.92 |
LongBench v2 | 50.50 | 48.67 |
SysBench | 67.88 | 67.22 |
GaRAGe | 42.61 | 40.73 |
RAGTruth | 77.59 | 77.88 |
RelyToolBench | 63.16 | 69.29 |
........
The quantized model's higher score on one test does not demonstrate that quantization universally improves reasoning. Sampling variation and harness behavior can affect individual outcomes.
Teams should test accuracy, tool-call validity, latency and peak memory on their target application and runtime rather than extrapolate from a single leaderboard.
··········
HARDWARE SUPPORT: RTX, HOPPER, BLACKWELL AND JETSON.
NVIDIA documentation identifies support across datacenter and local GPU platforms, including A100, H100/H200, GB200, GeForce RTX 5090, DGX Spark, RTX Spark and Jetson Orin and Thor.
These devices vary in available memory, bandwidth, compute capabilities and power requirements. Hardware compatibility does not imply equal speed or an identical maximum practical context at the same batch size.
NVIDIA describes PyTorch and vLLM deployment options and NVIDIA Dynamo for its preview serving path. Its release materials do not establish one independently measured tokens-per-second comparison across all listed devices.
........
Hardware | Examples | Context |
Ampere | A100 | Datacenter |
Hopper | H100, H200 | High throughput |
Blackwell | GB200 | Modern datacenter |
GeForce RTX | RTX 5090 | Local desktop |
Compact AI | DGX Spark, RTX Spark | Local systems |
Jetson | Orin, Thor | Edge deployment |
........
··········
DFLASH SPECULATIVE DECODING.
NVIDIA references a companion DFlash drafter compatible with the model for speculative decoding. The technique proposes tokens using an auxiliary mechanism and verifies them against the primary model, potentially reducing sequential generation cost.
Actual acceleration depends on draft acceptance rates, serving software, hardware and memory footprint. A drafter is an additional component, so its overhead matters on constrained local devices.
The model card does not establish a universal speedup factor for every application or hardware platform.
··········
LICENSE, COMMERCIAL USE AND OPEN WEIGHTS.
NVIDIA distributes the checkpoints publicly under OpenMDW License Agreement version 1.1 and indicates commercial and noncommercial use subject to applicable terms. The Google Gemma 4 foundation has its own licensing context.
The NVIDIA-distributed checkpoint should not be labeled simply Apache 2.0 merely because its base model has that association. Developers should review the full applicable license before deployment or redistribution.
Open weights do not imply that complete training data, reward specifications and intermediate states are public. NVIDIA's hosted preview API carries separate service terms from downloading and self-hosting weights.
··········
APPLICATIONS FOR LOCAL AGENTS.
For voice assistants, Bolt can process transcribed text, invoke tools and return text to a separate speech-synthesis component. Local placement can reduce network dependence, but privacy depends on the application's actual data flows.
Interactive game characters may benefit from the model's role consistency and multi-turn instruction following. Embedded interfaces can map natural-language commands to constrained local functions, provided the application controls permissions.
The BF16 checkpoint offers a starting point for domain-specific fine-tuning, while the quantized checkpoint emphasizes distribution and memory efficiency.
··········
LIMITATIONS, SAFETY AND BENCHMARK INTERPRETATION.
The model is text-only, its 64K context does not guarantee accurate retrieval throughout the whole window, and its vendor-reported scores have not been presented as independent reproductions. Several competitors outperform Bolt on particular long-context evaluations.
The quantized 4GB figure does not represent all inference resources. Agents still require external orchestration, permission boundaries and recovery logic. Training-data cutoff also limits factual freshness without retrieval.
Before production use, developers should evaluate prompt injection, unauthorized tool use, sensitive data exposure, hallucinations and multi-step failure modes on representative workloads.
··········
HOW BOLT E4B FITS INTO NVIDIA'S NEMOTRON STRATEGY.
NVIDIA combines a Gemma 4 foundation with agent-oriented reinforcement learning, configurable reasoning, compact quantized distribution and compatibility with its inference ecosystem.
Bolt E4B is designed for the trade-off between local resource constraints and reliable interactive behavior rather than leading every benchmark category. Its strongest positioning is in memory-aware assistants and tool-using applications.
The official evaluation results indicate meaningful conversational gains over the base model, alongside weaker results on certain long-context tasks. Developers can select BF16 for customization or Q4_K_M for local execution, after measuring real application success and hardware requirements.
··········
FOLLOW US FOR MORE.
·····
DATA STUDIOS
·····
datastudios.org
