What it is

NVIDIA released Nemotron 3.5 Bolt E4B on October 23, 2026 — a small language model post-trained from Google's Gemma 4 E4B. Rather than re-pretraining, NVIDIA used pure reinforcement learning (RLVR + RLHF, via NeMo-RL / NeMo-Gym) to tune Gemma 4 E4B into an agent base that "follows instructions, calls tools, and keeps answers short."

The E4B = Effective 4B in the name: the weight file actually carries ~7.5B parameters, but ~3.5B of those are embedding lookup tables — the "effective parameters" that actually compute per token are ~4B. So it's a "nominally 4B, skeleton 8B" model.

It supports 64K input + 64K output, text-in / text-out. It is not a cloud flagship chat model — its position is a lightweight on-device agent: voice assistants, game NPCs, robot / IoT front-ends — anywhere you need "chat plus tool calls inside a small VRAM budget."

Local deployment guide

The biggest trap first: on Ollama, running ollama run nemotron-3.5-bolt pulls the 25GB 74B model, not E4B. You must pin the small one explicitly:

ollama run nemotron-3.5-bolt:e4b

VRAM vs disk: "4GB" is the runtime VRAM footprint, not the file size. The official Q4_K_M GGUF is 6.18 GB, but only ~2.5GB of working weights sit on the GPU at runtime while the ~3.7GB embedding table stays in system RAM. Measured Q4_K_M VRAM on an RTX is 3.1–4.3 GB.

  • An 8GB-VRAM card is enough (RTX 3060 / 4060 / 4070 class) — near-zero cost.
  • No discrete GPU? Pure CPU + RAM works, with ≥8GB RAM (16GB recommended if you multitask).
  • Apple Silicon Macs supported: Ollama ships an MLX build; a 16GB Mac is comfortable.
  • Offline: once downloaded, Ollama / LM Studio / llama.cpp run fully air-gapped — conversations never leave your machine.

llama.cpp route: grab the Q4_K_M GGUF and start llama-server with --jinja for an OpenAI-compatible endpoint; add --spec-type draft-dflash to speed up (a 0.74GB DFlash speculative-drafting model proposes 7 tokens at a time).

Turn off "thinking": thinking is on by default; for crisp answers use ollama run nemotron-3.5-bolt:e4b --think=false or /set nothink inside Ollama. NVIDIA's recommended sampling is temperature 1.0 + top_p 0.95.

No local deploy? NVIDIA API Catalog serves nvidia/nemotron-3.5-bolt-e4b; grab a key from build.nvidia.com.

Same-tier comparison

NVIDIA's own target is the ~4B tier. The scores below are NVIDIA's NeMo Gym BF16 measurements — vendor self-test, treat as reference:

DimensionBolt E4BGemma 4 E4BQwen3.5-4BNano 4B
Params8B total / 4B effective8B / 4.5B effective4B3.97B
ArchGemma 4, RL post-trainedDense + per-layer embedGated DeltaNet + attn hybridMamba-2 + Transformer hybrid
Context64K128K262K262K
LicenseOpenMDW-1.1Apache 2.0Apache 2.0NVIDIA Open Model License
Inputtexttext / image / audiotext / image / videotext

Verdict: Bolt E4B wins on "follows instructions + multi-turn tool use" — IHEval (instruction-level adherence) 84.8, nearly 18 points above the closest same-tier Gemma 4 E4B (67.0); that's its headline selling point. But it is not built for coding; on long-context and function-call accuracy Qwen3.5-4B is actually stronger (RULER 64K 95.1 vs 87.0, Tau2-Bench 75.1 vs 55.1).

So which to pick: need a local agent that strictly follows the system prompt and calls tools inside a small VRAM budget → Bolt E4B is the default in this tier. Need long context, pure chat/writing, or code → the Gemma 4 E4B / Qwen3.5-4B you likely already have fit better, at similar speed and better conversation.

Sources

  • NVIDIA model card and build.nvidia.com (GA 2026-10-23)
  • Community write-ups: Ollama / LM Studio / llama.cpp deployment docs; media restatements of BF16 benchmarks (vendor self-test — scores per official model card)

← Back to all guides