Research Vault

NVIDIA Nemotron 3.5 Bolt E4B (BF16)

Properties1
tagshuggingface, text-generation, agentic, instruct, tool-calling, edge, slm, dense, gemma4-base, openmdw

5 min read

NVIDIA Nemotron 3.5 Bolt E4B (BF16)

One-line: NVIDIA’s Oct-23 edge SLM — an 8B/4B-effective post-train of Gemma-4-E4B-it under OpenMDW-1.1, built for local voice/device assistants and game-NPC agents; the card publishes a full agentic tool-use table measured under NVIDIA’s own NeMo Gym harness — BFCL-v3 68.51, τ²-Bench 55.10, ToolSandbox 68.77, When2Call Multi-Turn 63.36, IHEval 84.79. ⚠ no coding benchmarks published at all (no HumanEval/SWE-bench/LiveCodeBench/Aider) — evaluate it as a small tool-calling/instruction-following model, not a coding agent.

Release date: NVIDIA’s blog (Oct 15) announced Super VL + Bolt 74B-A9B as “available now, with Bolt E4B coming Oct. 23”, and the Oct-23 RTX Spark post repeats “Bolt E4B 4GB … releases Friday, Oct. 23”. The repo’s createdAt is 2026-10-05 (staged empty/private), while the commits API pins the actual content drop: initial commit 2026-10-23T15:12Z, then README updates 15:59Z and 17:12Z — i.e. 4.8h before this run. Per the “createdAt ≠ weight release date” rule, 2026-10-23 is the release date.

⚠ All numbers below are NVIDIA-reported, measured under their NeMo Gym harness (“they may differ from vendors’ self-reported numbers” — i.e. NVIDIA re-ran the competitors too, which is a stronger provenance than a pure vendor card, but it is not a third party). No independent (Artificial Analysis / BenchLM / LMArena) evaluation found. Ratings start at the .eval_results-style table on the card, not a third party.

Coding benchmarks

BenchmarkScoreSource
HumanEval—not published
HumanEval+—not published
MBPP—not published
LiveCodeBench—not published
Aider polyglot—not published
SWE-bench Verified—not published
SWE-bench Multilingual—not published

No coding evaluation of any kind is published for this model — its benchmark slate is tool use, instruction following, persona/RP and RAG-groundedness.

Agentic / tool-use benchmarks

BenchmarkScoreSource
BFCL-v368.51NVIDIA — NeMo Gym; vs Gemma 4 E4B-it 66.39, Qwen3.5 4B 70.95, Nemotron-3 Nano-4B 61.10
τ²-Bench (Tau2-Bench)55.10NVIDIA — NeMo Gym; best in the BF16 row (Gemma base 38.50, Qwen3.5 4B 75.06, Nano-4B 29.20)
ToolACE—not published (card uses ToolSandbox instead)
GAIA—not published
ToolSandbox68.77NVIDIA — NeMo Gym (Qwen3.5 4B 69.54)
When2Call (multi-turn)63.36NVIDIA — NeMo Gym; best in row (Gemma base 62.90, Nano-4B 27.70)
AgentIF25.81NVIDIA — NeMo Gym; best in row
MultiChallenge53.10NVIDIA — vs Gemma base 42.08, Qwen3.5 4B 51.76
RelyToolBench (tool hallucination)63.16NVIDIA — vs Gemma base 61.60
IHEval (instruction hierarchy)84.79NVIDIA — best in row (Gemma base 67.04)
StructEval-T80.40NVIDIA — vs Qwen3.5 4B 76.01
LongMemEval72.10NVIDIA — vs Gemma base 70.95

Reasoning (context only)

BenchmarkScoreSource
MMLU / MMLU-Pro—not published
GPQA Diamond—not published
MATH—not published
AIME 2025 / 2026—not published
IFBench63.79NVIDIA — vs Gemma base 42.84, Qwen3.5 4B 58.91
InverseIFEval46.84NVIDIA — best in row
RoleBench / RoleMRC67.05 / 80.63NVIDIA — persona/RP rows, best in row
EQ-Bench v339.40NVIDIA — vs Gemma base 41.17
SysBench67.88NVIDIA — best in row
RULER (64K)86.99NVIDIA — vs Qwen3.5 4B 95.15
LongBench v2 (48K)50.50NVIDIA — vs Qwen3.5 4B 59.93
GaRAGe / RAGTruth42.61 / 77.59NVIDIA — groundedness rows, best in row
Brevity Benchmark82.52NVIDIA — best in row

Efficiency

  • Active params: 4B effective / 8B total (dense Gemma-4-E4B backbone)
  • Context length: 64K input / 64K output
  • VRAM: BF16 ≈16 GB — the card’s quickstart is a single H100 (vllm serve … --tensor-parallel-size 1). The first-party Q4_K_M sibling nvidia/NVIDIA-Nemotron-3.5-Bolt-E4B-4GB is the ~4 GB llama.cpp path; a DFlash NVFP4 drafter (…-BF16-DFlash-NVFP4) ships for speculative decoding (7 draft tokens in the example config).
  • Runtime: stock transformers (model_type: gemma4); vLLM v0.31.0+ with --tool-call-parser gemma4 --reasoning-parser gemma4 --enable-auto-tool-choice; llama.cpp recipe linked as “pending” on the 4GB card. Chat-template controls enable_thinking plus a runtime thinking budget, and a separate conversational_mode. Recommended temp 1.0 / top_p 0.95. Card has a technical report + NVIDIA blog; post-training data cutoff May 2026. Text in / text out (the underlying Gemma 4 E4B is multimodal but this post-train is text-only).
  • License: OpenMDW-1.1 (permissive, NVIDIA’s own; license: other + license_name: openmdw-1.1, license_link points off-repo to openmdw.ai — a LICENSE fetch 404s, which is expected for this license, not a missing license).

What makes it notable

The third-generation NVIDIA ACE SLM tier, released 2026-10-23 as the small end of the Nemotron 3.5 Bolt family (Super VL and Bolt 74B-A9B shipped 10-15; this is the “E4B 4GB / embedded interfaces” tier). For this tracker the value is the tool-use provenance: the card publishes a 22-row agentic slate run under NVIDIA’s own NeMo Gym harness, scoring both the model and its baselines — τ²-Bench 55.10 (its base Gemma-4-E4B-it scores 38.50 in the same harness), MultiChallenge 53.10 (base 42.08), IHEval 84.79 (base 67.04), RoleBench 67.05 (base 49.95): consistent double-digit post-training gains well past the 5-pp bar. In-book caveat: on several rows the cheaper Qwen3.5-4B baseline is ahead of it (BFCL-v3 70.95, τ² 75.06, RULER 95.15, LongBench v2 59.93), so Bolt E4B’s pitch is not raw tool-calling accuracy but NVIDIA’s ACE/edge stack — reachy-companion demo, DGX Spark/RTX deployment, a 4 GB Q4_K_M build, and a first-party speculative drafter. And it is not a coding model: zero coding evals, and the card targets voice assistants, interactive personas and in-game NPCs (SkyrimNET). Included here because it is an open-weight, locally-runnable, tool-calling model with real published agentic numbers and a documented release inside the window — flagged so it is not mistaken for a coding-agent candidate.

See also

Graph View