What shipped
Bolt E4B is the smallest of the three Nemotron 3.5 models NVIDIA announced on 15 October. Super VL and the 74B Bolt shipped that day. E4B arrived on its announced date, 23 October, with GGUF and BF16 checkpoints, a DFlash drafter for each, and listings across the usual local-inference tools.
The pitch is specific. NVIDIA positions Bolt E4B primarily for conversational agents rather than coding-specialist workloads. NVIDIA built it for agents that talk: voice and device assistants, game characters that hold a persona while calling tools, and robot or industrial interfaces. It adds a conversational mode that favors short, speech-ready replies, configurable reasoning with a runtime thinking budget, and multi-turn tool use in a 64K context. Support covers 35 languages, built on a base model pre-trained on more than 140.
The training choice is worth noting. There is no supervised fine-tuning stage in the recipe. NVIDIA post-trained Gemma 4 E4B with reinforcement learning alone, running every sample in both reasoning and non-reasoning modes. NVIDIA has published training datasets and a post-training recipe alongside the weights, giving researchers a basis to inspect and attempt to reproduce the training approach. The release does not establish that independent reproduction will produce identical results.
Where it leads
On the capabilities closest to the pitch, the lead is wide. IHEval, which tests whether a model respects instruction hierarchy when instructions conflict, puts Bolt E4B at 84.8, against 67.0 for its own base model and 60.6 for Qwen3.5-4B. RoleMRC, a test of holding a character's identity, scores 80.6 against Qwen's 65.8. MultiChallenge, which tracks consistency across turns, scores 53.1, just ahead of Qwen's 51.8 and well ahead of Gemma's 42.1.
Counted across the full table, Bolt E4B holds the top score in 15 of 21 benchmarks. NVIDIA also claims up to 80% fewer output tokens than peers in reasoning-enabled comparisons, which matters for a voice agent: fewer tokens means a shorter wait and a lower decoding bill.
Two cautions apply to all of it. Every score was measured by NVIDIA, in NVIDIA's NeMo Gym harness, against a comparison set NVIDIA chose. And two of the 21 benchmarks, Brevity Bench and When2Call-MT, are NVIDIA's own evaluations that it says will be released later. The table is evidence, but it is the vendor's evidence.
The row that matches the job
Tau2-Bench simulates multi-turn customer tasks: a user with a problem, an agent with tools, several turns to resolve it. NVIDIA's own description in the table is exactly that. For customer-service-style agents that must resolve tasks over several turns, Tau2-Bench is one of the most directly relevant tests in NVIDIA's table.
There, Bolt E4B scores 55.1. Qwen3.5-4B scores 75.1. The gap in the Q4_K_M build, the one most people will download, is 52.3 against 61.7. Qwen also leads on BFCL-v3 tool calling, ToolSandbox, RULER at 64K and LongBench v2, so the pattern is not one stray row. Qwen leads on several of NVIDIA's reported tool-use and long-context evaluations, although not every measure in those categories. On persona, brevity and instruction control, Bolt is ahead.
That split is the practical finding. A game character that must stay in role, follow a system prompt and call a crafting tool is a different job from an assistant that must work through a billing dispute. NVIDIA's table, read whole, supports the first claim better than the second.
The 4 GB label, and the wait on small hardware
The 4GB in the name is a memory envelope from NVIDIA's own llama.cpp tests, not a fixed property. On an RTX 5070 the Q4_K_M build used 3.5 GB with a short input, rising to 4.3 GB with a 32K cache. On Jetson boards it used 4.2 to 5.0 GB. The 4GB label describes a target runtime memory envelope, not a constant VRAM requirement or the download size. NVIDIA's measurements vary with hardware and context configuration; its Hugging Face listing separately shows a roughly 6.18 GB quantized artifact. Long contexts will push past the label; short ones sit under it, potentially leaving memory for other pipeline components, depending on hardware and configuration.
Speed is where the target hardware shows. NVIDIA measured 548 ms to first token and 133.4 tokens per second on an RTX 5070 with a 3K-token input. That is a more promising latency profile than its Jetson Orin Nano result, but end-to-end voice responsiveness also depends on speech recognition, speech synthesis, buffering and the application pipeline. A Jetson Orin Nano, the 8 GB board in NVIDIA's own robot demo, took 5,785 ms to its first token and generated 15.5 tokens per second. That is a demonstration configuration, not a responsive assistant. If the plan is a Jetson-based product, the latency table deserves more attention than the benchmark table.
Sources
- [1] NVIDIA research post, "Introducing NVIDIA Nemotron 3.5 Bolt E4B 4GB" - primary source: benchmark tables, VRAM and time-to-first-token tables, training method and availability list, read in fullRead source
- [2] Hugging Face model card, nvidia/NVIDIA-Nemotron-3.5-Bolt-E4B-4GB - release date and GA 10/23/2026; BF16 and Q4_K_M benchmark table measured by NVIDIA under the NeMo Gym harness; license OpenMDW-1.1; Q4_K_M GGUF file size 6,180,446,208 bytes in the repository file listingRead source
- [3] NVIDIA blog, "Open From Data Center to Desktop: NVIDIA Expands Nemotron Family With Highest Efficiency Super VL and Bolt Models" (15 Oct announcement; E4B dated Oct 23)Read source
- [4] MarkTechPost release write-up, 23 Oct 2026 (timing cross-check only)Read source