CuteGuy☣️ (@cutetoxicguy) on X

What it claims

Nemotron 3.5 Bolt E4B is an NVIDIA open-weights model (8B total, 4B active parameters per token) for local agents in about 4GB of GPU memory. NVIDIA built it on Google's Gemma 4 E4B, unlike the larger Nemotron 3.5 Bolt 74B. It targets conversational AI, interactive game characters and embedded devices. Weights are released under OpenMDW 1.1, which permits commercial use. The source is a third-party X post by @cutetoxicguy restating NVIDIA's announcement, not NVIDIA's own page.

How it works

NVIDIA post-trained Gemma 4 E4B with NeMo RL and NeMo Gym, targeting instruction following, structured outputs, tool use and conversation. Reasoning is configurable: thinking can be turned on or off, reasoning tokens can be capped, and the model can switch to short conversational replies for voice assistants.

Inference can use DFlash speculative decoding. A separate draft model predicts several tokens, and Bolt verifies them in parallel. NVIDIA reports up to 21% higher throughput than Gemma 4's original MTP drafter. The post does not give the setting for that figure.

Evidence

NVIDIA ran the Q4_K_M quantized model in llama.cpp with about 3K input tokens and 50 output tokens:

• RTX 5070 (12GB): 133.4 tokens/sec, 548ms to first token, 3.5GB peak VRAM.
• Jetson Orin Nano (8GB): 15.5 tokens/sec, 5.8s to first token, 4.2GB peak memory.

NVIDIA's BF16 evaluations against Qwen3.5-4B show a split result:

• IHEval: 84.8 vs 60.6 (Bolt ahead).
• RoleBench: 67.0 vs 53.7 (Bolt ahead).
• Tau2-Bench: 55.1 vs 75.1 (Qwen ahead).
• RULER 64K: 87.0 vs 95.1 (Qwen ahead).

NVIDIA also reports about 40% higher average benchmark scores than Nemotron 3 Nano 4B, and up to 80% fewer output tokens in selected comparisons. The post does not list which benchmarks or comparisons.

Demos include a Minecraft blacksmith NPC that negotiates payment, checks recipes, crafts items and hands them to players. Another demo runs Bolt with Cosmos 3 Edge on a Jetson-powered Reachy Mini robot.

How to use it

GGUF and BF16 checkpoints are on Hugging Face (link truncated in the post, under huggingface.co/nvidia/). Supported runtimes are llama.cpp, Ollama, LM Studio, vLLM and SGLang. The memory figures above are for Q4_K_M. The post gives no run commands or exact model IDs.

Limits

All benchmarks are NVIDIA's own, not independent. The speed figures use a short 3K-token prompt and 50-token output; the post states that longer contexts raise memory use and latency, so the 64K window will not fit the quoted 3.5GB to 4.2GB. Qwen3.5-4B leads on Tau2-Bench, an agentic tool-use benchmark, by 20 points, and on RULER 64K long-context retrieval, which matters for multi-step agent workloads. Time to first token on Jetson Orin Nano is 5.8 seconds even at 3K input. Benchmarks were run in BF16, while the speed tests used Q4_K_M, so quantized accuracy is not reported.