Nemotron 3.5 Bolt E4B

Nemotron 3.5 Bolt E4B (full name NVIDIA Nemotron 3.5 Bolt E4B 4GB) is an open small language model that NVIDIA released on October 23, 2026 for fast conversational agents: voice and device assistants, interactive game characters, and robot or IoT interfaces running locally on GeForce RTX GPUs, Jetson modules, RTX Spark and DGX Spark.^[1][3] It is not a new architecture. NVIDIA built it by post-training Google's Gemma 4 E4B instruction-tuned model with reinforcement learning, and the model card lists 8 billion total parameters, 4 billion of them "effective".^[2][3] The "4GB" in the name refers to the memory target of the 4-bit quantized release in NVIDIA's target workloads, not to the size of the download.^[1][7]

NVIDIA first announced the model on October 13, 2026 as the third-generation small language model for NVIDIA ACE, its technology suite for in-game AI characters, and said players could try it in the Skyrim modding platform SkyrimNET from October 23.^[10] NVIDIA announced it in the same post as the much larger Nemotron 3.5 Bolt 74B-A9B, which was released on October 15, 2026 and built by a different method: compression of Nemotron 3.5 Super VL.^[10][23] The weights, a speculative decoding drafter, and the training recipe are published under the OpenMDW-1.1 license.^[1][3]

Overview

Sources: model cards, configuration file and Hugging Face API metadata, accessed October 25, 2026.^[2][3][7][8]

Name and base model

"E4B" is Google's size label for the base model. Google's Gemma 4 model card says the "E" in E2B and E4B "stands for 'effective' parameters": the two smallest Gemma 4 models use Per-Layer Embeddings (PLE), which give each decoder layer its own small embedding for every token. Those tables are large but serve only as lookups, so Google counts the E4B model as 4.5 billion effective parameters, or 8 billion including embeddings.^[9] NVIDIA's cards round the same figures to "8B (4B effective)".^[2][3] The "4GB" suffix appears only on NVIDIA's model: the research release says the model is "designed to pack conversational agent intelligence into as little as ~4GB in its target workloads".^[1]

Bolt E4B keeps the Gemma 4 E4B network unchanged. NVIDIA's NeMo AutoModel documentation describes it as keeping "the Gemma 4 E4B architecture" while refreshing "the weights and the chat template", and the BF16 checkpoint loads as Gemma4ForConditionalGeneration in Hugging Face Transformers.^[8][15] The released configuration has a hidden size of 2,560, a 262,144-token vocabulary and tied input and output embeddings. It sets the maximum position count to 65,536, half of the 128K context that Google lists for Gemma 4 E4B.^[8][9]

The base model is multimodal. Google lists text, image and audio input for Gemma 4 E4B, with a vision encoder of about 150 million parameters and an audio encoder of about 300 million.^[9] NVIDIA's model card describes Bolt E4B as text in, text out, but the BF16 configuration keeps the Gemma 4 audio and vision encoder settings, and NVIDIA's technical report evaluates both audio and vision input (see the multimodal results below).^[2][3][8]

Announcement and NVIDIA ACE

On October 13, 2026, in a blog post about software for the RTX Spark laptop launch, NVIDIA said Nemotron 3.5 Bolt E4B 4GB would be "the third-generation small language model (SLM) for NVIDIA ACE AI game characters" and would release on Friday, October 23. The post described ACE as "the suite of technologies for building in-game AI agents". It said the first-generation model had been in a public beta earlier in 2026 for PUBG, where players could team up with AI teammates, and that the second-generation model had previewed in Aniimo at Gamescom, turning creatures into voice-responsive companions.^[10]

The same post named SkyrimNET, an AI non-player-character modding platform by MinLL for The Elder Scrolls V: Skyrim, as the first place to try the model from October 23. NVIDIA said SkyrimNET gives each NPC "personal memory, voice and autonomy" and supports both cloud and local inference across several models.^[10] NVIDIA's release page shows the model running a blacksmith trader NPC that checks the player's payment, then calls tools to look up a crafting recipe, gather ingredients, craft the item and hand it over.^[1]

On October 23, NVIDIA's @NVIDIAAI account announced the model as "Our newest small open model", built "for local assistants, interactive characters and robots that can carry on a conversation and use tools to get things done".^[11] The accompanying video showed Reachy Mini, an open-source desktop robot from Pollen Robotics and Hugging Face, chatting and moving, with Bolt handling conversation and NVIDIA Cosmos handling vision. NVIDIA said both models ran locally on a Jetson Orin Nano 8GB.^[11][22] The research release identifies the vision model as Cosmos 3 Edge, which supplies real-time visual understanding to Bolt; Bolt then calls scripts from the robot's animation libraries and produces its spoken replies.^[1]

Intended uses

NVIDIA groups the target applications into three categories:^[1]

NVIDIA argues that leaderboard rank is a poor guide for these workloads. Its release page says "a model that excels at agentic coding is often not the best fit for a fast home assistant", and lists the traits it sought instead: conversing naturally, deciding whether and when to interleave dialogue with tool calls, honoring persona and output constraints, interpreting supplied application or world state, and grounding replies in that context as it grows.^[1] It describes typical target workloads as single sessions (batch size 1) with 2,500 to 16,000 input tokens and fewer than 500 output tokens.^[1]

Features and chat template

NVIDIA extended the Gemma 4 chat template rather than replacing it. When an assistant turn contains both text and tool calls, the template places the text before the calls and their results, so that the assistant can acknowledge a request or explain an action before acting. The existing enable_thinking and preserve_thinking controls are kept.^[2]

The conversational-mode instruction shipped in the GGUF chat template reads: "You are having a spoken conversation with the user. Respond as briefly as possible while still addressing their questions fully." It is appended to an existing system prompt, or becomes the system prompt if there is none.^[2][7] NVIDIA chose this wording using results from its internal brevity benchmark.^[2]

The model cards also carry prompting guidance aimed at game and robot agents. They advise defining the agent and the user as separate entities with their own possessions and knowledge, using labels such as YOUR INVENTORY or SHARED CHEST, treating an inaccessible inventory as unknown rather than empty, and sending state updates as separate user or system messages in the correct order relative to tool results.^[3]

Training

Bolt E4B was post-trained with reinforcement learning only. NVIDIA used reinforcement learning with verifiable rewards (RLVR) and from human feedback (RLHF) across several environments, run with NVIDIA's NeMo RL and NeMo Gym software.^[1][2] The model card names the algorithm as GRPO and describes an asynchronous RL setup that decouples training from inference.^[3]

Source: technical report, Table 1.^[2]

The technical report gives these training choices:^[2]

• Batch settings. A global batch size of 256 with 8 rollouts per prompt, and maximum generation lengths of 4,096 tokens for reasoning samples and 1,024 for non-reasoning samples.
• Non-reasoning focus. Every training sample was used in both reasoning and non-reasoning modes, with rewards profiled separately for each mode and a 2:1 ratio of non-reasoning to reasoning samples in each batch. NVIDIA says the aim was to emphasize accuracy without explicit reasoning.
• Verbosity control. Reasoning traces and answers grew substantially longer as RL progressed. NVIDIA countered this by raising group-relative length penalties in the GenRM environment and disabling overlong filtering in the other environments, and found that cycling the penalties and filtering on and off during training improved accuracy while keeping outputs short.
• No supervised fine-tuning stage. NVIDIA tried supervised fine-tuning on Nemotron post-training data first. It improved multi-turn tool calling and let later RL reach a higher ceiling in fewer steps, but training on data from diverse teacher models degraded Gemma's conversational quality and increased verbosity, so the released model skipped that stage.

The report says post-training used English data only.^[2] The model card's dataset section instead lists English, French, German, Italian, Japanese, Spanish and Chinese as the primary post-training languages, so NVIDIA's two documents differ on this point.^[2][3]

Quantization and memory

NVIDIA converted the selected BF16 checkpoint to the Q4_K_M format with llama.cpp, using importance-matrix (imatrix) calibration on about 1 million tokens of RL rollouts taken from the same training step as the checkpoint. NVIDIA reports that, on average accuracy, this imatrix Q4_K_M build beat a plain Q4_K_M conversion and imatrix IQ4_XS and IQ4_NL builds.^[1][2] The single GGUF file in the 4GB repository is 6,180,446,208 bytes, about 6.2GB.^[7]

NVIDIA's measured memory use stays between 3.1GB and 5.0GB across the configurations it published. All runs used Q4_K_M at batch size 1 on llama.cpp (CUDA), measured with AIPerf; "cache" is previously cached context and ISL and OSL are new input and output tokens:^[1]

NVIDIA notes that results vary with hardware, inference software and configuration.^[1] Unsloth's guide likewise says the ~4GB figure is VRAM, "not total system RAM or download size", and suggests the model runs in roughly 4GB of VRAM or 6GB of system RAM.^[13]

Evaluations

NVIDIA evaluated the model on 21 benchmarks in five groups, against Gemma-4-E4B-it, Alibaba's Qwen3.5 4B and NVIDIA's own earlier Nemotron 3 Nano 4B. All results are NVIDIA's own measurements in its NeMo Gym harness, and the cards warn that they "may differ from vendors' self-reported numbers".^[2][3] Most benchmarks ran with reasoning enabled; RoleBench, EQ-Bench v3 and RULER ran with reasoning off. Many scores depend on LLM judges or simulated users (GPT-4o, GPT-4.1, GPT-5.1, GPT-5.2, GPT-5.5, Qwen3-235B-A22B-Instruct-2507, or an ensemble of Claude Sonnet 5.5, Gemini 4 Argon and GPT-6.1 Sol), and some benchmarks were modified: RoleBench was run as a custom pairwise comparison against Gemma-4-E4B-it, InverseIFEval and SysBench were restricted to or translated into English, and the brevity benchmark and When2Call-MT are internal benchmarks that NVIDIA says it will release later.^[2]

BF16 results with reasoning mode as specified per benchmark (best score in bold):^[2]

By NVIDIA's count, Bolt E4B scored highest on 15 of the 21 benchmarks: four of five conversation tests, one of four tool-use tests, all five instruction-control tests, two of four context tests and all three grounding tests. With reasoning disabled for every model it led 14 of 21.^[2] Its clearest gaps were to Qwen3.5 4B on Tau2-Bench (55.10 against 75.06), RULER and LongBench v2; Gemma 4 E4B kept a small lead on EQ-Bench v3.^[2] The research release summarizes the results as roughly 40% higher average scores than Nano 4B and roughly 10% higher than the other peers evaluated.^[1]

NVIDIA also claims the gains come with shorter outputs: the technical report says Bolt generated fewer output tokens on average than Gemma-4-E4B-it and Qwen3.5 4B in both reasoning and non-reasoning modes, and the release page puts the reduction at up to 80% in reasoning-enabled comparisons.^[1][2] In a headline chart averaging six benchmarks for Q4_K_M models on an RTX 5090 (about 3,000 input and 50 output tokens), NVIDIA plotted cloud reference points of 77.1 for Gemini 3.8 Flash at a 2,449.5 ms time to first token and 76.2 for Claude Sonnet 5.5 at 3,002.2 ms.^[1][2]

The quantized Q4_K_M build held up in NVIDIA's tests and scored above the BF16 checkpoint on several benchmarks, including MultiChallenge (55.60), RoleBench (71.21), ToolSandbox (74.13) and RelyToolBench (69.29). In that comparison NVIDIA used Unsloth's Q4_K_M GGUF files for Gemma 4 E4B and Qwen3.5 4B.^[2]

Multilingual and multimodal results

Because post-training used English only, NVIDIA checked whether the base model's other abilities survived:^[2]

NVIDIA concludes that translation quality was preserved more consistently than multilingual knowledge and reasoning. It reports that audio understanding improved even though training was text-only, while vision showed "slight but consistent degradation", which it attributes to misalignment between the vision encoder and the retrained language model. Training the vision encoder alone brought no meaningful improvement, and NVIDIA lists restoring vision quality as future work.^[2] Both Bolt E4B and Gemma 4 E4B accept speech input of up to 30 seconds per turn. Because 23.36% of AudioMC user turns exceed that limit, NVIDIA notes that its AudioMC results do not reflect the complete conversations.^[2]

DFlash drafter and inference speed

The release includes a DFlash drafter for speculative decoding, a method based on the block-diffusion drafting paper by Jian Chen and colleagues.^[2][19] The drafter is a five-layer Transformer (five Qwen3-style decoder layers) with 360 million parameters, excluding the token embeddings and output head it reuses from Bolt. It reads hidden states from five intermediate layers of the target model, attends to a sliding window of the most recent 1,024 positions, and proposes seven tokens per step in one non-causal forward pass. Bolt verifies them in one pass and keeps the longest valid prefix, so the target model's output distribution is preserved.^[2][5] NVIDIA trained the drafter in its Model Optimizer toolkit by distilling Bolt's next-token distribution on 3.79 million conversations, using prompts from Nemotron-Post-Training-Dataset-v2 with responses regenerated by Bolt in both reasoning modes.^[2]

NVIDIA also tested LiLiCorr, the correction-head extension of DFlash used by the 74B Bolt model. LiLiCorr raised acceptance length by about 5% for E4B, but because each decoding step of a model this small is short, the extra head's overhead left per-user speed about the same, so NVIDIA kept plain DFlash.^[2]

Sources: drafter model cards.^[5][6]

NVFP4 quantization of the drafter cut the weights read per drafting step from 2.06GB to 0.58GB, with at most a 2% loss in acceptance length; on DGX Spark the default W4A16 drafter gave a 1.13x per-user speedup over the BF16 drafter at concurrency 1.^[2] The NVFP4 drafter's mean acceptance length, the average number of tokens produced per verification step out of a maximum of eight, was 4.01 on MT-Bench and 3.88 on the SPEED-Bench qualitative split with greedy decoding, falling to 3.65 and 3.50 with sampling at temperature 1.^[2][6] On MT-Bench, acceptance was highest for math, extraction and coding prompts and lowest for humanities, roleplay and writing.^[6]

NVIDIA's measured speedups from the drafter, in the BF16 NIM build with reasoning on and greedy decoding:^[2]

With reasoning off, the speedup fell to between 2.32x and 3.11x. Time to first token changed by less than 100 ms in every setting, because speculative decoding shortens decoding rather than prefill.^[2] For local 4-bit use, NVIDIA compared Q4_K_M Bolt with DFlash against Q4_K_M Gemma-4-E4B-it with Google's multi-token-prediction (MTP) drafter on the coding subset of SPEED-Bench at batch size 1. Bolt reached 272.6 tokens per second against 225.9 on an RTX 5070 (1.21x) and 124.4 against 118.4 on an RTX Spark N1X (1.05x). The "up to 21% higher throughput" claim on NVIDIA's release page matches the RTX 5070 result.^[1][2]

Availability and deployment

NVIDIA published four checkpoints on Hugging Face. Hugging Face metadata shows the repositories were created on October 5, 2026 (October 9 for the NVFP4 drafter), and the model cards give October 23 as the release date.^[3][7]

NVIDIA also packaged the BF16 model as an NVIDIA NIM microservice built on vLLM 0.30.0, published for amd64 and arm64, with a free API endpoint at build.nvidia.com. The NIM enables the DFlash drafter by default with a draft length of seven, uses a 65,536-token context, accepts text, image and audio input, and exposes OpenAI-compatible Chat Completions, Completions and Responses endpoints plus an Anthropic-compatible Messages endpoint. NVIDIA says it validated the NIM on DGX Spark (GB10), RTX PRO 6000 and RTX PRO 4500 Blackwell Server Edition, and H100 80GB GPUs, and measured no throughput overhead against a standalone vLLM server.^[2]

NVIDIA lists Hugging Face, Ollama, LM Studio, Unsloth, Canonical, Jetson AI Lab and the NVIDIA In-Game Inferencing SDK as ways to get the model.^[1] The In-Game Inferencing SDK is NVIDIA's toolkit for running local AI models in-process inside games and applications.^[20] Several of these are documented separately:

• Ollama added the model as the e4b tag of its nemotron-3.5-bolt library entry, alongside the 74B model, describing it as an 8-billion-parameter model with 4 billion effective that runs on Windows, Linux and Apple Silicon. As of October 25, 2026 its default e4b tag is listed at 7.2GB to 7.9GB with a 128K context window and text and image input, with further Q4_K_M, Q8_0 (9.8GB) and BF16 (17GB) tags at 128K and three Apple MLX tags at 64K; the 128K figures differ from NVIDIA's 64K context, and the sizes from its ~4GB memory target.^[12]
• Unsloth documents a 64K context for the model and publishes its own GGUF repository, based on NVIDIA's BF16 checkpoint, with quantizations from 2-bit (UD-IQ2_M) up to BF16, vision projector files and a Q8_0 DFlash drafter.^[13][16]
• Jetson AI Lab lists the model as text-only on Jetson, at 8B total and 4B effective parameters with a 64K context and the Q4_K_M GGUF, and lists Orin and Thor modules from the Jetson Orin Nano 8GB up to the Jetson AGX Thor developer kit; its guide serves text chat with llama.cpp.^[14]
• NeMo AutoModel provides an eight-GPU full-parameter fine-tuning recipe and a two-GPU LoRA recipe for the BF16 checkpoint, both run on H100 80GB GPUs with the MedPix-VQA dataset and with the vision and audio towers frozen.^[15]

Relationship to other Nemotron models

NVIDIA calls Nemotron 3 Nano 4B "our prior model" when stating Bolt E4B's 40% improvement.^[1] Unlike Nano 4B, a member of the Nemotron 3 family with a hybrid Mamba-Transformer design, Bolt E4B inherits Google's Gemma 4 E4B network, and its model card notes the Apache 2.0 terms of the Gemma 4 base alongside the OpenMDW-1.1 license.^[3][21]

Reception

Coverage on release day relayed NVIDIA's figures. MarkTechPost described the model as "agent behavior in about 4GB of VRAM" that leaves room for speech and vision models, and noted the RL-only training and the benchmarks where Qwen3.5 4B and Gemma 4 E4B kept leads.^[17] AlphaSignal argued that the ~4GB target matters on an 8GB Jetson or a 12GB consumer GPU, where the language model shares memory with the rest of an agent stack, but observed that the Jetson Orin Nano's time to first token of almost 5.8 seconds "can disrupt a voice interface". It wrote that Qwen3.5 4B's leads on Tau2-Bench, BFCL-v3 and LongBench v2 favor Qwen for structured customer-service flows, function-heavy applications and long-document reasoning.^[18]

References

1. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12 ^13 ^14 ^15 ^16 ^17 ^18 ^19 ^20 ^21 ^22NVIDIA. NVIDIA Nemotron 3.5 Bolt E4B 4GB. NVIDIA Nemotron research release, October 23, 2026. Accessed October 25, 2026.
2. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12 ^13 ^14 ^15 ^16 ^17 ^18 ^19 ^20 ^21 ^22 ^23 ^24 ^25 ^26 ^27 ^28 ^29 ^30 ^31 ^32 ^33 ^34 ^35 ^36NVIDIA. Nemotron 3.5 Bolt E4B 4GB: An Open and Efficient Small Language Model for Fast Conversational Agents. Technical report, October 23, 2026. Sections 1-9; Tables 1-12.
3. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12 ^13 ^14 ^15 ^16 ^17NVIDIA. NVIDIA-Nemotron-3.5-Bolt-E4B-4GB model card. Hugging Face, released October 23, 2026. Accessed October 25, 2026.
4. ^NVIDIA. NVIDIA-Nemotron-3.5-Bolt-E4B-BF16 model card. Hugging Face, released October 23, 2026. Accessed October 25, 2026.
5. ^1 ^2NVIDIA. NVIDIA-Nemotron-3.5-Bolt-E4B-4GB-DFlash model card. Hugging Face, released October 23, 2026. Accessed October 25, 2026.
6. ^1 ^2 ^3NVIDIA. NVIDIA-Nemotron-3.5-Bolt-E4B-BF16-DFlash-NVFP4 model card. Hugging Face, released October 23, 2026. Accessed October 25, 2026.
7. ^1 ^2 ^3 ^4 ^5Hugging Face. Hub API metadata for nvidia/NVIDIA-Nemotron-3.5-Bolt-E4B-4GB and companion repositories (creation dates, file sizes, GGUF chat template, parameter counts). Accessed October 25, 2026.
8. ^1 ^2 ^3 ^4NVIDIA. NVIDIA-Nemotron-3.5-Bolt-E4B-BF16 configuration. `config.json`, text, vision and audio settings. Accessed October 25, 2026.
9. ^1 ^2 ^3Google DeepMind. Gemma 4 E4B IT model card. Hugging Face. Accessed October 25, 2026.
10. ^1 ^2 ^3 ^4 ^5Gerardo Delgado, NVIDIA. All Systems Go: Global PC App Ecosystem Ready for NVIDIA RTX Spark. NVIDIA Blog, October 13, 2026.
11. ^1 ^2NVIDIA AI (@NVIDIAAI). Post announcing Nemotron 3.5 Bolt E4B 4GB. X, October 23, 2026.
12. ^Ollama. nemotron-3.5-bolt and tags. Ollama model library. Accessed October 25, 2026.
13. ^1 ^2Unsloth. NVIDIA Nemotron 3.5 Bolt: Run Locally Guide. Accessed October 25, 2026.
14. ^NVIDIA. Nemotron 3.5 Bolt E4B 4GB. Jetson AI Lab. Accessed October 25, 2026.
15. ^1 ^2NVIDIA. NVIDIA-Nemotron-3.5-Bolt-E4B-4GB-BF16. NeMo AutoModel model-coverage documentation, GitHub. Accessed October 25, 2026.
16. ^Unsloth. NVIDIA-Nemotron-3.5-Bolt-E4B-GGUF. Hugging Face model repository, created October 23, 2026. Accessed October 25, 2026.
17. ^Asif Razzaq. NVIDIA Releases Nemotron 3.5 Bolt E4B 4GB: An Open Small Language Model for Local Voice Agents, Game Characters and Robots. MarkTechPost, October 23, 2026.
18. ^AlphaSignal Newsroom. NVIDIA's Nemotron 3.5 Bolt Squeezes a Local AI Agent Into 4GB. AlphaSignal, October 23, 2026.
19. ^Jian Chen, Yesheng Liang, and Zhijian Liu. DFlash: Block Diffusion for Flash Speculative Decoding. arXiv:2602.06036, 2026.
20. ^NVIDIA. NVIDIA In-Game Inferencing SDK. NVIDIA Developer. Accessed October 25, 2026.
21. ^1 ^2NVIDIA. Nemotron 3 Nano 4B: A Compact Hybrid Model for Efficient Local AI. Hugging Face community article, March 17, 2026.
22. ^Pollen Robotics and Hugging Face. Reachy Mini: The Open-Source Robot for Today's and Tomorrow's AI Builders. Hugging Face blog. Accessed October 25, 2026.
23. ^1 ^2NVIDIA. NVIDIA Nemotron 3.5 Bolt 74B-A9B 24GB. NVIDIA Nemotron research release, October 15, 2026. Accessed October 25, 2026.