NVIDIA Nemotron 3.5 Bolt Is Called 4GB; Its 4-Bit File Is 6.18 GB
A.I. / news
NVIDIA Nemotron 3.5 Bolt Is Called 4GB; Its 4-Bit File Is 6.18 GB
The Gemma 4 E4B derivative wins seven of eight benchmark rows against its base on NVIDIA's own harness, but loses five of eight to Qwen3.5 4B.

NVIDIA's Nemotron 3.5 Bolt E4B 4GB, an 8-billion-parameter model it lists as 4 billion effective, is a post-trained version of Google's Gemma 4 E4B, and its model card shows the 4GB in the name is not the file size. The Q4_K_M GGUF is 6.18 GB.
The card on Hugging Face gives a release date of October 23, 2026. NVIDIA says it is built for "fast local voice and device assistants, interactive personas, and embedded interfaces."

What the card says about the model
The model is released under the OpenMDW License Agreement, version 1.1. The card says the underlying Gemma 4 model is governed by the Apache License 2.0, and that Bolt is "ready for commercial or non‑commercial use." The card shows no access-request gate on the weights themselves, though some datasets it references are gated.
Training data is listed as more than 20 trillion tokens, "Hybrid: Automated, Manually-Collected, Synthetic," covering 2013 to December 2025, with a May 2026 post-training cutoff.
Context is 64K tokens in and 64K out. Google's own Gemma 4 E4B card lists 128K and describes the model as 4.5 billion effective parameters, 8 billion with embeddings. NVIDIA's card says 4 billion effective, so the two disagree on the figure by half a billion.
The benchmark table is NVIDIA's
Every score below is NVIDIA-measured at Q4_K_M, mostly in reasoning mode, with NVIDIA's NeMo Gym harness. The card says they may differ from other vendors' self-reported numbers. No independent party is named as having run them.
| Benchmark | Bolt | Gemma 4 E4B-it | Qwen3.5 4B |
|---|---|---|---|
| MultiChallenge | 55.60 | 39.36 | 52.23 |
| RoleBench | 71.21 | 60.51 | 52.35 |
| IHEval | 82.40 | 66.61 | 62.87 |
| Tau2-Bench | 52.32 | 38.72 | 61.74 |
| RULER (64K) | 85.76 | 86.14 | 95.37 |
| LongMemEval | 69.92 | 67.95 | 76.60 |
Against its own base, Bolt scores higher on seven of the eight rows on the card and lower on RULER at 64K, 85.76 against 86.14. Against Qwen3.5 4B it wins three rows (MultiChallenge, RoleBench, IHEval) and loses five, including Tau2-Bench by 9.4 points and RULER by 9.6.
- Qwen3.5 4B61.74 points
- Nemotron 3.5 Bolt52.32 points
- Gemma 4 E4B-it38.72 points
Source: NVIDIA model card on Hugging Face, accessed 2026-11-20
On a Google card, Gemma 4 E4B-it lists 58.6% on GPQA Diamond and 69.4% on MMLU Pro. The Bolt card does not report those two, so the post-training's effect on general knowledge is not on the page.
Hardware and what is missing
NVIDIA lists Ampere, Hopper and Blackwell GPUs, Linux and PyTorch as supported, and says it tested on hardware ranging from 1 to 8 H100 or H200 cards to a GeForce RTX 5090, DGX Spark and Jetson Orin and Thor. The card gives no phone or laptop numbers, and no latency or tokens-per-second figures for the edge parts it names.
The card does not explain the "4GB" label. The 6.18 GB file at 4-bit is the only size it gives.
Another open-weight release with a disputed card is covered in our Tencent Hy4-Flash report, and Google's closed-model standing is in our Gemini 4 Neon ranking report. The number to watch is an independent run of the same eight benchmarks, which the card does not mention.
Sources
More in A.I.
- 01Chrome Built a Fake Web to Test Auto Browse Against Prompt InjectionTwo Chrome AI Security engineers describe a gVisor sandbox that serves generated sites at real URLs, with injections chosen by a request header.
- 02Anthropic Prices IPO at $83, and Two Outlets Disagree on the ProceedsFortune reports about $79.4 billion raised on 980 million shares; The Next Web, citing Bloomberg and the FT, reports at least $68 billion on 801.4 million.
- 03Tencent Hy4-flash Card Says Weights Are Out; Launch Coverage Says PlannedThe Hugging Face page lists 220B parameters and a custom licence, while launch coverage gives 224B and a future release, at $0.070 per million input tokens.
- 04Anthropic Says Claude Closed a Black Hole Bound Left Between 8.503 and 12The company's post names one model for the result; the paper's own record names two others, and no outside physicist has reviewed it publicly.