# Fix list: NVIDIA Nemotron 3.5 Bolt E4B 4GB From Anchor Terminal's listing at https://www.anchorterminal.com/tools/nemotron-3-5-bolt-e4b-4gb, the October 2026 research run, assessed 24 October 2026. Grade C, 61 out of 100. This is everything the published grade says the listing lacks, the biggest possible gain to the total first. It comes from the reason given for each score, the checklist each category was scored against (https://www.anchorterminal.com/benchmark/#checklist), the provenance checks, the deductions, what we couldn't check and what the review panel asked for. A fix counts at the next check, once it's public. For a coding agent working on NVIDIA Nemotron 3.5 Bolt E4B 4GB: work through the items below in the product, its docs and its public pages. Each category gives the reason for its score, with the points each checklist item earned, and the checklist itself, so the gap is the items that earned less than their points. Change the product, not the wording, and keep a note of what you changed and where it's published. ## 1. Risk, 80 out of 100, up to 6.2 more on the total Why it scored 80: File safety earns 25, safetensors and GGUF only, and the documented routes run no repository code (an undocumented Inference Endpoints handler.py sits in the BF16 repository). Provenance earns 15, NVIDIA's linked organisation with a hash for every file. Safety evaluation earns 0, since NVIDIA names its content-safety and bias datasets but publishes no results. Agent guidance earns 15, for instruction-hierarchy training NVIDIA documents and a subcard on prompt injection and the safeguards to add. Licence risk keeps all 15 under OpenMDW-1.1 and Apache 2.0. The security programme earns 10 for the card's reporting route and NVIDIA's public bulletins. The checklist (https://www.anchorterminal.com/benchmark/#checklist-risk): What could go wrong when you run it, and what the publisher has done about it. A higher score means less risk. - 0 to 25, file safety. 25 for weights in safetensors or GGUF only, 15 when pickle files (.bin, .pt, .pth) sit beside them, 0 for pickle only. 10 off when the documented way to load it or to build its prompts runs code from the repository (`trust_remote_code`, or the publisher's scripts in place of a template). - 0 to 15, provenance. 10 when the files come from the publisher's own organisation, the one its site or docs link to, and 5 for a checksum on every file. - 0 to 20, safety evaluation. 20 when results are published and someone other than the publisher ran them or can rerun them, 10 for the publisher's own results, 0 for none. - 0 to 15, guidance for agent use. 15 for documented prompt-injection guidance or an instruction hierarchy the model was trained to follow, 10 for a statement of what its trained refusals cover and the safeguards to add around it, 5 for general cautions, 0 for none. - 0 to 15, licence risk. 15 to start, 5 off for each of these, a use policy or field-of-use limits beyond the law, terms the licensor can change or revoke, a revenue or user threshold, and a licence that changed without a dated note or whose link is broken. Not below 0. - 0 to 10, a security programme. 5 for a way to report a flaw (security.txt, a disclosure policy or a security contact), 5 for advisories handled in public. When our reference-machine runs start, measured robustness (prompt injection with AgentDojo, over-refusal with XSTest) joins Risk. ## 2. Documentation and upkeep, 65 out of 100, up to 5.4 more on the total Why it scored 65: The model card earns 20, all seven items with limitations in the linked subcard. Training data earns 15, the RL datasets published and a cut-off given, but the card's data section describes a larger SFT corpus the report says wasn't used, and the base's pre-training is described only in general. Last change earns 20 for 23 October 2026. Responsiveness earns 0, as nobody has asked anything yet. Version policy earns 5 as a first release whose GGUF was replaced on release day with only a commit title. File health earns 5, since the card and the template disagree on the thinking default and both cards carry leftover lines. The checklist (https://www.anchorterminal.com/benchmark/#checklist-upkeep): - 0 to 20, the model card. Intended use, limitations, architecture, parameters, context length, languages and knowledge cut-off. 20 for all seven, 3 off for each that's missing. - 0 to 20, training data. 20 for the data published or listed with sources and sizes, 15 for a data card that names the sources or the mix, the cut-off and the filtering, 10 for a general description with a cut-off, 5 without one, 0 for none. - 0 to 20, the last change to the files, the template or the reference code. 20 within 30 days, 15 within 90, 10 within 180, 5 within a year, 0 older. - 0 to 20, responsiveness. The publisher's replies to the 40 most recent questions and reports in the repository's discussions or issues. 20 for most, 10 for some, 5 for a few, 0 for none. - 0 to 10, version policy. 10 for earlier versions kept in their own repositories with a dated note when one is replaced, 5 for earlier versions kept without notes or a first release, 0 when files were replaced in place without a note. - 0 to 10, file health. 10 when config, tokenizer, template and card agree and no report of a broken file waits without a reply, 5 when one does, 0 when the files contradict the card. ## 3. Reach, 90 out of 100, up to 3.8 more on the total Why it scored 90: Memory earns 30, since NVIDIA's Q4_K_M GGUF is 6.18 GB, 7.4 GB with a fifth added. Runtimes earn the full 25 from released llama.cpp, vLLM, Transformers, SGLang, Ollama and LM Studio. Builds earn 10, as NVIDIA's only quantised build is the Q4_K_M GGUF and the DFlash repositories are drafters. Terms earn 10, because OpenMDW-1.1 is permissive with no threshold or use policy but isn't an OSI licence, and the Gemma 4 base adds plain Apache 2.0. Download earns 10, with no gate. Hosted earns 5, as NVIDIA's own API serves it. The checklist (https://www.anchorterminal.com/benchmark/#checklist-reach): Whether an agent can run it, and where. - 0 to 30, memory. The smallest build from the publisher or a runtime's own library that keeps 4-bit precision or better, at its file size plus a fifth for the runtime and a working context. 30 for 16 GB or less, 25 for 24 GB, 20 for 32 GB, 15 for 64 GB, 10 for what fits our reference machine (the GPU share of 128 GB of unified memory), 5 for 256 GB, 0 above. - 0 to 25, runtimes. 5 for each of llama.cpp, Ollama, vLLM, SGLang, LM Studio and Hugging Face Transformers that runs the architecture in a released version, from the runtime's own code, library or documentation. A publisher's fork or an unreleased branch doesn't count. 25 at most. - 0 to 15, builds from the publisher. 15 for quantised builds at two precisions or more (FP8, 4-bit, GGUF), 10 for one, 0 when only other people's quantisations exist. - 0 to 15, terms of use. 15 for an OSI-approved licence with nothing added, 10 for a custom licence that lets anyone use it commercially for free below a threshold or under a use policy, 5 for research or non-commercial use only, 0 for no licence or unclear terms. - 0 to 10, download. 10 when the files download with no account and no terms to accept, 5 for a click-through that needs an account, 0 when a person approves each request. - 5, hosted too. The publisher's own API serves it, or at least two inference providers do, so an agent without the hardware can still call it. When our reference-machine runs start, speed on that machine (tokens per second and time to first token at the build that fits) joins Reach. ## 4. Agent interface, 85 out of 100, up to 2.3 more on the total Why it scored 85: The chat template earns 20, a Jinja file in the repository and the GGUF. The tool-calling format earns 25, with OpenAI-style tools, the gemma4 parser named and an asynchronous pattern documented. Trained tool calling earns 20, multi-step tool-use training and several calls per turn. Structured output earns 5, trained but with no instructions. Reasoning control earns 10, thinking per request, the reasoning returned apart and a budget. Context earns 5 for 64K, half the base's 128K. The checklist (https://www.anchorterminal.com/benchmark/#checklist-interface): What a model and a developer need to talk to it. Speech, image and other models with no chat interface read the tool-calling and structured-output lines for what the model takes and returns (voices, timestamps, streaming, typed outputs), and the note says how. - 0 to 20, the chat template. 20 for a template in the repository (Jinja, in the tokenizer config or its own file), 10 when the prompt format is only given as code you run. - 0 to 25, the tool-calling format. 25 when the publisher documents how tools are declared and how calls come back, with the parser to use in a named runtime, 10 when only the template or a runtime's own recipe says it. - 0 to 20, tool calling the model was trained for. 20 with several calls in one turn, 10 for one call at a time. - 0 to 15, structured output. 15 when JSON or JSON Schema output is documented for the model with how to ask for it, 5 when the card claims it without saying how. - 10, reasoning control. Thinking that can be turned off or budgeted per request, with where the reasoning text comes back. - 0 to 10, context. 10 for 128K tokens or more trained, 5 for 32K to 127K. Context reached only by scaling the position encoding counts one band down. ## What we couldn't check What we couldn't read counted as absent. Publishing it on a page a plain HTTP fetch can read (not only in a browser) lets the next check count it. - Capability is computed later. The independent results found on 24 October 2026, Artificial Analysis IFBench 0.611 and Scale's MultiChallenge 56.03, both read as instruction following, so with one area Capability would stay pending and the total capped at 61. If MultiChallenge counts in another area it wouldn't - Scale's small-model MultiChallenge tab was added on 22 October 2026, a day before release, so Scale had the model early. Whether NVIDIA submitted it isn't stated; it's recorded as Scale's own run - Terms read OpenMDW-1.1 as a custom licence at 10, as for Lightning, Ultra, Parse and the Laguna models. Apache 2.0 alone would be 15 - File health is 5. A stricter reading of the card contradicting the template on the thinking default gives 0, Upkeep 60 - Version policy is 5. The GGUF was replaced in place on release day to fix its context metadata, with a commit title as the only note; reading that as no note gives 0, Upkeep 60 - Training data is 15. NVIDIA's own post-training data is published in full, which could read as 20, but the card's data section describes a corpus the report says wasn't used - SGLang is credited from its released gemma4_mm code. Nobody documents Bolt on SGLang; the cap is reached without it - sameModel stays empty. NVIDIA's API serves nvidia/nemotron-3.5-bolt-e4b, but no listing covers that endpoint - unchecked, www.nvidia.com/.well-known/security.txt, left as unknown from the earlier Nemotron listings ## Weaknesses - 64K tokens of context against the base's 128K, and Ollama's and LM Studio's pages show 128K anyway - No safety evaluation results published - One quantised build from NVIDIA, and its GGUF is text only with no vision or audio projector - The card says thinking is on by default while the template defaults to off - Independent results so far cover instruction following only (IFBench and MultiChallenge) ## What costs an agent a turn today The notes we give agents before they call it. Each one is a workaround an agent shouldn't need. - 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 - 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 ` to cap thinking - 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 - 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 - Use `temperature=1.0` and `top_p=0.95` in both modes, as the card says - Set `conversational_mode` true for spoken replies, which appends a brevity instruction to the system prompt - 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 - 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 ## What the review panel asked for - publish safety evaluation results - release notes for replaced files - align card with template - Document JSON output - Fix the thinking default text - Publish more quantised builds - State the hosted price ## When it's done Send what changed and where it's published as a dispute (https://www.anchorterminal.com/builders/#disputes, or `POST https://www.anchorterminal.com/api/v1/contact` with `"kind": "dispute"`). Disputes are answered in public, and the listing is checked again by the same checklist. Paying for an audit or a listing claim changes nothing here.