NVIDIA Nemotron 3.5 Bolt E4B: 4GB Agent Model Trained With Reinforcement Learning Only
NVIDIA released Nemotron 3.5 Bolt E4B on October 23, a small language model built specifically for local deployment in interactive agent applications. The model targets voice assistants, game non-player characters, and robot interfaces running on RTX, Jetson, and Spark hardware, where cloud round-trips are impractical.
At Q4_KM quantisation, the model uses 3.1 to 4.3 GB of VRAM. On an RTX 5070, Bolt with DFlash speculative decoding reaches 272.6 tokens per second — 1.21x faster than its base model, Gemma 4 E4B, at the same precision.
Architecture and Training
The model has 8 billion total parameters with 4 billion effective, and a 64,000-token context window. NVIDIA post-trained Gemma 4 E4B using reinforcement learning only, skipping supervised fine-tuning entirely. The training pipeline used RLVR and RLHF through NeMo-RL and NeMo-Gym, with length penalties applied to discourage verbose outputs unsuitable for spoken interfaces.
Developers get two runtime controls: a conversational style toggle that shortens responses for speech synthesis, and a reasoning toggle that trades latency for accuracy on harder tasks. The combination is designed for interactive loops where a game NPC may need rapid responses but an agentic workflow step may need more deliberation.
Benchmark Results
NVIDIA’s in-house comparison ran Bolt against Gemma 4 E4B, Qwen3.5-4B, and Nemotron 3 Nano 4B across 21 benchmarks. Bolt led in 15 of 21.
| Benchmark | Bolt E4B | Gemma 4 E4B |
|---|---|---|
| IHEval | 84.8 | 67.0 |
| BFCL-v3 (tool selection) | 67.3% | — |
| MultiChallenge (Scale Labs) | 56.03% | — |
Scale Labs independently ranked Bolt first among small models on MultiChallenge. On BFCL-v3, a tool selection and argument completion test, Bolt scored 67.3% against Qwen 3.5 4B’s 70.6% at the same quantisation level — Qwen leads on that specific task despite trailing on aggregate quality measures.
NVIDIA’s comparison used an RTX 5090 for benchmark measurements. The hardware is vendor-supplied, and these are not third-party evaluations.
Why the RL-Only Training Matters
Skipping supervised fine-tuning is a deliberate architecture choice, not a cost-cutting measure. SFT tends to push models toward longer, more hedged responses that cover expected training distributions. For a conversational interface — especially a game NPC that must respond in under 200 milliseconds — verbosity is a latency problem. RL-only post-training lets the model optimise directly against human preference signals and task completion, with length penalties shaping response style without a prior SFT baseline pulling it toward verbosity.
The same approach underpins larger reasoning-focused models. At 4B effective parameters, Bolt is testing whether RL-only training transfers into the edge-deployment tier.
Deployment Targets
The model is designed for RTX laptops and desktops, NVIDIA Jetson edge boards, and Spark embedded systems. Applications envisioned in the model card include always-on voice assistants, game characters that require persistent dialogue state, and robot controllers that call external tools while managing real-time interaction. All three workloads share a common constraint: they cannot wait for a cloud API round-trip.
Bolt ships under an open model licence. NVIDIA has not disclosed pricing beyond the base hardware dependency.