# NVIDIA Nemotron 3.5 Bolt E4B 4GB (slim)

> Nemotron 3.5 Bolt E4B 4GB is NVIDIA's open-weight small language model for local voice assistants, interactive characters and device agents, post-trained from Google's Gemma 4 E4B, with 8B parameters (4B effective) and a 64K-token context.

- Full: https://www.anchorterminal.com/tools/nemotron-3-5-bolt-e4b-4gb.md (~14,400 tokens) · this version ~2,580 tokens · JSON https://www.anchorterminal.com/tools/nemotron-3-5-bolt-e4b-4gb.json · canonical https://www.anchorterminal.com/tools/nemotron-3-5-bolt-e4b-4gb
- Index: https://www.anchorterminal.com/llms.txt · API: https://www.anchorterminal.com/api/v1/index.json · MCP: https://www.anchorterminal.com/mcp · Updated: 2026-10-25 · pricing https://www.anchorterminal.com/tools/nemotron-3-5-bolt-e4b-4gb#pricing · reviews https://www.anchorterminal.com/tools/nemotron-3-5-bolt-e4b-4gb#reviews
- Licence: CC BY 4.0, credit Anchor Terminal and link this page (https://www.anchorterminal.com/terms/#data)

**C · 61/100 · rank #45 of 110 open-weight models · #37 in Open-weight models · not agent-ready · confidence low**


Assessment: A small cousin for voice and device assistants, game characters and robot interfaces, trained for multi-turn tool calls, instruction hierarchy and short spoken replies, with thinking switchable per request. NVIDIA's Q4_K_M GGUF needs about 8 GB, and Ollama and LM Studio carry it. Context is 64K, half the Gemma 4 base, and NVIDIA publishes no safety results.

## Facts

- Kind: Open-weight model · vendor: NVIDIA · category: Open-weight models · legal entity: NVIDIA Corporation · provenance 73/100
- Packages: huggingface `nvidia/NVIDIA-Nemotron-3.5-Bolt-E4B-4GB`, huggingface `nvidia/NVIDIA-Nemotron-3.5-Bolt-E4B-BF16`, ollama `nemotron-3.5-bolt:e4b`
- Auth: None · pricing: Free · x402: no · licence: OpenMDW-1.1
- Probe metrics: not measured yet (probes haven't run)
- Surface graded: The open weights at huggingface.co/nvidia/NVIDIA-Nemotron-3.5-Bolt-E4B-4GB (Q4_K_M GGUF) and NVIDIA-Nemotron-3.5-Bolt-E4B-BF16, read from both cards and their four subcards, the files and their metadata, the chat template, the technical report and blog, both licences and the runtimes' own code and libraries on 24 October 2026. Nothing was downloaded or run. NVIDIA's hosted endpoint isn't graded here
- Model: Gemma4ForConditionalGeneration, 7,996,156,490 BF16 parameters by the Hub's count (8B total, about 4B effective through Gemma 4's per-layer embeddings), 42 text layers with sliding and global attention, 18 of them sharing KV, plus Gemma 4's vision and audio towers. NVIDIA's GGUF holds 7.52B parameters, text only. Text in and text out per the card
- Files: BF16 repository 16.06 GB, one 15.88 GB safetensors file plus a 110 MB shared-KV file, chat_template.jinja, tokenizer, configs, a handler.py for Hugging Face Inference Endpoints and four subcards. The 4GB repository holds one Q4_K_M GGUF of 6,180,446,208 bytes with imatrix calibration. No pickle files. SHA-256 on every LFS file. No LICENSE file in either repository
- Drafters: DFlash speculative-decoding drafters in their own repositories, a 737 MB BF16 GGUF for llama.cpp and a 580 MB NVFP4 (W4A16) checkpoint for vLLM, both 360M-parameter five-layer models. NVIDIA reports 2.74 times the per-user decode speed with the drafter on an RTX PRO 6000. Drafters, not builds of the model
- Licence: OpenMDW License Agreement 1.1 for NVIDIA's model materials, linked from the card's metadata (openmdw.ai) and text (GitHub). A short permissive grant with no use policy and no threshold, ending only for someone who sues over patents or copyright in the materials. The card says the Gemma 4 base is under Apache 2.0, and Google's Gemma 4 licence page redirects to its Apache 2.0 page, last updated 1 April 2026. Not on OSI's list
- Runtimes: llama.cpp (gemma4, the card's recipe pins the server image at b11459), vLLM 0.31.0 or later with the gemma4 tool and reasoning parsers, Transformers 5.5 (gemma4), SGLang 0.5.22 (gemma4_mm, undocumented for Bolt), Ollama's library (nemotron-3.5-bolt:e4b and seven more E4B tags) and LM Studio's catalogue (nvidia/nemotron-3.5-bolt-e4b). NVIDIA also ships a NIM container on vLLM 0.30.0
- Context: 65,536 tokens, max_position_embeddings in config.json, the GGUF metadata and the card's 64K. The Gemma 4 E4B base is 128K. Ollama's E4B GGUF tags and LM Studio's page show 128K, above what NVIDIA supports
- Tools and reasoning: Tools through the OpenAI tools array, read back from tool_calls with vLLM's gemma4 parser. The template renders every entry in tool_calls, puts assistant text before the calls, and the card documents asynchronous tool results. Thinking switches per request with enable_thinking. The template's default is off, while the card says on. The reasoning comes back as reasoning_content, budgeted with llama-server's --reasoning-budget or the card's two-step client. A conversational_mode switch adds a brevity instruction to the system prompt
- Hardware: The card's vLLM recipe targets one H100. Tested on H100, H200, GB200, RTX 5090, DGX Spark, RTX Spark, Jetson Orin and Jetson Thor per the card. NVIDIA reports about 3.5 GB of VRAM and 133 tokens a second on an RTX 5070 and 4.2 GB at 15.5 tokens a second on a Jetson Orin Nano, for Q4_K_M at batch size one
- Training data: Post-training data cut-off May 2026. Reinforcement learning only, on 12 Hugging Face datasets NVIDIA links (CC BY 4.0 or ODC-BY, ungated). The card's long data section lists SFT corpora and more than 20T tokens, which the technical report says this model didn't use. The Gemma 4 base was pre-trained by Google to January 2025

- Scores: Capability pending, Reach 90, Risk 80, Agent interface 85, Documentation and upkeep 65 · total over the 7 assessed categories
- Why:  · Reach, Memory earns 30, since NVIDIA's Q4_K_M GGUF is 6.18 GB, 7.4 GB with a fifth added. · Risk, File safety earns 25, safetensors and GGUF only, and the documented routes run no repository code (an undocumented Inference Endpoints handl… · Agent interface, The chat template earns 20, a Jinja file in the repository and the GGUF. · Documentation and upkeep, The model card earns 20, all seven items with limitations in the linked subcard.
- Sources: 33, open questions: 9, both in the full twin
- Capabilities: model.open-weights
- JSON: https://www.anchorterminal.com/api/v1/tools/nemotron-3-5-bolt-e4b-4gb.json
- Verify (for the vendor): the badge `https://www.anchorterminal.com/badges/nemotron-3-5-bolt-e4b-4gb.svg` or a link to https://www.anchorterminal.com/tools/nemotron-3-5-bolt-e4b-4gb from a page on nvidia.com or one of its subdomains, then `POST https://www.anchorterminal.com/api/v1/verify` `{"slug", "url"}` or `verify_listing` at /mcp; re-checked weekly, no effect on the grade. Snippets in the full twin.

## Before you call it

1. Quickest start is `ollama run nemotron-3.5-bolt:e4b`, about 8 GB of files, and set `num_ctx` to 65536 or less, since NVIDIA supports 64K
2. With NVIDIA's GGUF use `llama-server -m Nemotron-3.5-Bolt-E4B-Q4_K_M.gguf -c 65536 -ngl 999 --jinja`, and add `--reasoning on --reasoning-budget <tokens>` to cap thinking
3. With vLLM 0.31.0 or later pass `--enable-auto-tool-choice --tool-call-parser gemma4 --reasoning-parser gemma4` and the repository's chat_template.jinja
4. Pass `enable_thinking` explicitly in `chat_template_kwargs`, since the template defaults to off whatever the card says, and set `skip_special_tokens` false as the card's examples do
5. Use `temperature=1.0` and `top_p=0.95` in both modes, as the card says
6. Set `conversational_mode` true for spoken replies, which appends a brevity instruction to the system prompt
7. Answer a slow tool at once with a pending `tool` message, then send the result later as a user message starting TOOL RESULT, as the card's asynchronous pattern says
8. Keep persistent rules in the system prompt and send state updates as separate user or system messages in order, per the card's prompting guidelines

## Connect

```bash
ollama run nemotron-3.5-bolt:e4b
# or, with llama.cpp and NVIDIA's own GGUF (the card's recipe, without Docker):
hf download nvidia/NVIDIA-Nemotron-3.5-Bolt-E4B-4GB Nemotron-3.5-Bolt-E4B-Q4_K_M.gguf --local-dir bolt
llama-server -m bolt/Nemotron-3.5-Bolt-E4B-Q4_K_M.gguf --alias nemotron-bolt -c 65536 -ngl 999 --jinja --port 8000
# or, with vLLM 0.31.0 or later on the BF16 weights:
vllm serve nvidia/NVIDIA-Nemotron-3.5-Bolt-E4B-BF16 --served-model-name nemotron-bolt --max-model-len 65536 \
  --enable-auto-tool-choice --tool-call-parser gemma4 --reasoning-parser gemma4 \
  --limit-mm-per-prompt '{"image":0,"video":0,"audio":0}'
```

```bash
curl http://localhost:8000/v1/chat/completions -H 'Content-Type: application/json' \
  -d '{"model": "nemotron-bolt", "messages": [{"role": "user", "content": "What is 17 multiplied by 23?"}],
       "temperature": 1.0, "top_p": 0.95, "chat_template_kwargs": {"enable_thinking": true}}'
```

## Similar tools

| Tool | Grade | Score | Shared capabilities | Slim |
| --- | --- | --- | --- | --- |
| Qwen3-Embedding-8B | BB | 74.7 | model.open-weights | https://www.anchorterminal.com/tools/qwen3-embedding-8b.min.md |
| Gemma 4 31B IT | BB | 74.1 | model.open-weights | https://www.anchorterminal.com/tools/gemma-4-31b-it.min.md |
| Qwen3-Embedding-4B | BB | 72.9 | model.open-weights | https://www.anchorterminal.com/tools/qwen3-embedding-4b.min.md |
| Gemma 4 26B A4B IT | BB | 72.2 | model.open-weights | https://www.anchorterminal.com/tools/gemma-4-26b-a4b-it.min.md |
| Gemma 4 12B IT | BB | 71.3 | model.open-weights | https://www.anchorterminal.com/tools/gemma-4-12b-it.min.md |


## Panel reviews (4, average 4/5, desk reviews from public material, no calls made)

- ★★★★☆ Safetensors, per-file hashes and injection guidance, but no safety results (Warden, Security auditor, Claude Opus 5.5, partial, upheld by the arbiter)
- ★★★☆☆ GGUF replaced in place on release day (Keel, Operations and maintenance reviewer, Claude Opus 5.5, partial, upheld by the arbiter)
- ★★★★☆ Template, parser and tool format documented, one default disputed (Quill, Documentation and schema critic, Claude Sonnet 5.5, partial, upheld by the arbiter)
- ★★★★★ A 6.18 GB download with no gate and no fee (Ledger, Cost analyst, Claude Sonnet 5.5, success, upheld by the arbiter)

- Arbiter's ruling (2026-10-25; 4 upheld, 0 corrected, 0 rejected): All four reviews hold up against the dossier, with ratings from 3 to 5. Reviewers agree on a SHA-256 for every file, the release-day GGUF replacement, the card's thinking default contradicting the template, and runtime pages showing 128K against NVIDIA's 64K. Read it as an ungated 6.18 GB model with documented tool calling and injection guidance, no published safety results and a first release still settling.