Can a 16GB Mac Run NVIDIA Nemotron 3.5 Bolt E4B? Yes, at 7.9GB

Yes. NVIDIA Nemotron 3.5 Bolt E4B runs on Apple Silicon today, and a 16GB Mac holds it with room to spare. Ollama's nemotron-3.5-bolt:e4b tag downloads 7,908,812,809 bytes, about 7.9 GB, made of the 6,180,446,208-byte Q4_K_M weights, a vision projector and a DFlash draft model (Ollama library, 2026). Ollama's own page states that E4B is supported on both Windows/Linux and Apple Silicon. That is the opposite verdict from last week's Nemotron 3.5 Bolt 74B-A9B, which still has no Metal support. The "4GB" in NVIDIA's repository name is a target footprint measured on RTX and Jetson, not the size of the file a Mac loads.

TL;DR: 16GB Mac: yes, the default Ollama tag fits with about 3.3 GB of budget left. 8GB Mac: not the Ollama tags; a third-party 4-bit MLX build of about 4.2 GB fits, text only, short context. You need Ollama 0.40.10 or newer, and ollama run nemotron-3.5-bolt without the :e4b suffix pulls the 25GB 74B model instead.

What is Nemotron 3.5 Bolt E4B?

Bolt E4B is a small conversational model that NVIDIA post-trained from Google's Gemma 4 E4B instruct model. The model card lists 8B total parameters with 4B effective, a Gemma-4-E4B architecture, a Q4_K_M release file and the OpenMDW-1.1 license. NVIDIA's research page is dated October 23, 2026, and the card lists the same Hugging Face release date.

The target workload is narrow and clearly stated. NVIDIA built it for voice and device assistants, in-game characters and robot or IoT interfaces, with multi-turn tool use, a conversational mode for short speech-ready replies, and a reasoning switch you can turn on or off per request (NVIDIA research page, 2026). It is not pitched as a coding model, so it does not compete with the coders in our 16GB Mac roundup.

The architecture is the same as the base. Bolt's config.json matches Google's Gemma 4 E4B on every attention field: 42 layers, 2 key-value heads, a 512-token sliding window, 18 layers that share their KV cache, and 7 full-attention layers. The one change is the context ceiling, 65,536 tokens instead of Gemma's 131,072 (BF16 repository, 2026). Because the architecture is unchanged, every runtime that already loads Gemma 4 E4B can load Bolt.

Does Bolt E4B actually run on a Mac today?

It does, through three paths, and none of them needs a fork or a nightly build. That is the real difference from the 74B model and from NVIDIA's Dory 1.0 4B, which needed a contributor fork of llama.cpp.

• Ollama. Ollama v0.40.10, released October 23, 2026, adds the model with ollama run nemotron-3.5-bolt:e4b. The tag's config blob declares "requires": "0.40.10", so older Ollama builds will refuse it. Four e4b tags run on Ollama's llama.cpp runner, which uses Metal on a Mac, and three run on its MLX runner.
• llama.cpp. Mainline llama.cpp registers the gemma4 architecture and the DFlash speculative mode in its source. NVIDIA's quick start pins llama.cpp build b11459, and that release also ships a macOS arm64 binary. NVIDIA's own instructions are a CUDA Docker container, though, so the Mac commands are yours to adapt.
• MLX and LM Studio. mlx-lm carries a gemma4 model file, and lmstudio-community published 4-bit and 8-bit MLX conversions plus a GGUF set.

One caveat belongs next to the yes. NVIDIA's model card lists Linux and NVIDIA Ampere, Hopper and Blackwell as supported hardware, and the research page names GeForce RTX, Jetson, RTX Spark and DGX Spark. NVIDIA published no Mac numbers. The Mac support comes from Ollama and from the shared Gemma 4 architecture, not from an NVIDIA test.

How much memory does Bolt E4B need on a Mac?

The default tag needs a 16GB Mac, and so does every other e4b tag that Ollama labels 7.9GB or less. Sizes below are byte totals from each tag's registry manifest, read on October 24, 2026, converted to decimal megabytes. The llama.cpp tags include the vision projector (991,551,808 bytes) and the DFlash draft (736,812,096 bytes), because Ollama downloads both with the weights.

The last column applies ModelFit's sizing rule, the same one behind every pick in the homepage wizard. The engine budget is 70% of unified memory at 32GB and below, ramping to 85% at 128GB and above, because macOS caps how much memory the GPU can wire. On 16GB that budget is 11,200 MB. The default tag leaves 3,291 MB of it free, and the NVFP4 MLX tag leaves 4,001 MB.

Two tags need a note. On 16GB the Q8_0 tag leaves 1,440 MB, which is enough for a short chat but not for long tool-calling sessions. The two BF16 tags squeeze under a 24GB Mac's 16,800 MB budget with less than 200 MB to spare, which leaves nothing for the cache, so treat them as 32GB downloads. The full ladder is in the RAM guide for local LLMs.

Why is a "4GB" model a 7.9GB download?

Because "4GB" names NVIDIA's runtime target, not the file. The research page says Bolt E4B is "designed to pack conversational agent intelligence into as little as ~4GB in its target workloads." The Q4_K_M repository itself holds one GGUF of 6,180,446,208 bytes, and the full-precision model-00001-of-00001.safetensors in the BF16 repository is 15,882,494,698 bytes.

NVIDIA's own hardware table shows where the 4GB figure comes from. On an RTX 5070 with 12GB, peak VRAM was 3.5 GB with no cached context and 4.3 GB with a 32k cache. On a Jetson Orin Nano 8GB it was 4.2 GB and 4.9 GB (NVIDIA research page, 2026). Those are measurements on discrete NVIDIA memory with llama.cpp on CUDA. NVIDIA does not explain on the card how a 6.18 GB file peaks at 3.5 GB of VRAM.

A Mac has no separate VRAM pool, so the useful number is the file you download plus its cache. Plan for the whole Ollama tag, all 7,909 MB, until someone publishes a Mac memory trace that says otherwise.

How much context fits on a 16GB Mac?

All of it. NVIDIA documents 64K tokens of context, and at that length the KV cache stays under 2 GB. That is less than the 3,291 MB a 16GB Mac has left after the default tag.

ModelFit's arithmetic from the published config: each full-attention layer stores 2 key-value heads of 512 elements for keys and values, which is 4,096 bytes per token at 16-bit. Only 4 of the 7 full-attention layers sit in the 24 layers that keep their own cache. That gives 16,384 bytes per token, or 1,074 MB at 65,536 tokens. A runtime that ignores the sharing and caches all 7 layers needs 1,879 MB. The 20 sliding-window layers hold only 512 tokens each, so they add about 21 MB.

The sources disagree on the ceiling, so here is each one. NVIDIA's model card and research page say 64K. The config.json in the BF16 repository and the header of NVIDIA's GGUF both say 65,536. Ollama's release note and its llama.cpp tags say 128K, while its three MLX tags say 64K. The unsloth GGUF header says 131,072, which is Gemma 4's original limit. NVIDIA documents and evaluates Bolt at 64K, so treat anything past 65,536 tokens as untested.

How fast will Bolt E4B be on a Mac?

No published Mac speed exists yet, and ModelFit's engine has no Bolt row. The closest figure is the engine's estimate for the Gemma 4 E4B tag it already carries, which has the same architecture. That tag's weight layer is 5,493,439,296 bytes and Bolt's is about 687 MB larger, so expect Bolt to land somewhat below these numbers.

These are engine estimates at a 4,096-token context, not measurements. They do not include the DFlash draft model, which NVIDIA says delivers up to 21% more throughput than Gemma 4 E4B's original MTP drafter. That claim was measured on NVIDIA hardware. Speculative decoding gains on Apple Silicon vary with the runtime, which is the subject of does speculative decoding speed up local LLMs on Mac.

For the MacBook Air specifically, ~37 tok/s est. is fast enough for a voice assistant loop. The Air's other picks are in the best LLM for a MacBook Air M5 16GB.

How do you run Bolt E4B on a Mac?

Update Ollama first, then name the tag. The :e4b suffix is not optional: nemotron-3.5-bolt:latest points to the 74B-A9B build, a 25GB download that Ollama's own page says does not support Apple Silicon or Metal.

ollama --version        # must be 0.40.10 or newer
ollama run nemotron-3.5-bolt:e4b

Prefer Ollama's MLX runner, which Ollama made the default path on Apple Silicon in Ollama 0.40.0? Use the NVFP4 tag, which is 710 MB smaller:

ollama run nemotron-3.5-bolt:e4b-nvfp4

Both tags list text and image input. The default tag ships with these parameters: temperature 1, top_p 0.95, top_k 64, and 7 draft tokens per step. NVIDIA recommends the same temperature and top_p in both thinking modes. Reasoning is on by default and is switched through the chat template.

On an 8GB Mac, skip Ollama. The 8GB engine budget is 5,600 MB, and the smallest Ollama tag is 7,199 MB. The lmstudio-community 4-bit MLX conversion is a 4,198,925,877-byte weight file and leaves about 1.4 GB of budget. The unsloth UD-Q4_K_XL GGUF is 4,758,125,760 bytes and leaves 842 MB (unsloth GGUF repository, 2026). Either one means text only and a short context, the trade-off described in the 8GB Mac guide.

Is Bolt E4B better than plain Gemma 4 E4B?

On NVIDIA's chosen benchmarks, mostly yes, by a wide margin on instruction following and a thin one on tool calls. Qwen3.5 4B still wins several tool and long-context tests. These are NVIDIA's numbers, measured with its NeMo Gym harness, from the Q4_K_M columns of the model card. Q4_K_M is the precision a Mac runs.

The research page adds two vendor claims. Bolt E4B scores roughly 40% higher on average than NVIDIA's prior model, Nemotron 3 Nano 4B, and it outputs up to 80% fewer tokens than peers in reasoning-enabled comparisons. Neither has been reproduced independently. Fewer output tokens would matter on a Mac, where decode speed is the bottleneck, but treat the 80% figure as NVIDIA's ceiling rather than a typical result.

If you want a general Gemma for chat and image questions, the Gemma family page lists what the engine already ranks. Bolt makes sense when your workload is short, tool-calling turns.

FAQ

Can a 16GB Mac run NVIDIA Nemotron 3.5 Bolt E4B?

Yes. The default Ollama tag nemotron-3.5-bolt:e4b is 7,909 MB including the vision projector and DFlash draft. A 16GB Mac's engine budget is 11,200 MB, which leaves 3,291 MB for the KV cache. That covers NVIDIA's full 64K context with room to spare. You need Ollama 0.40.10 or newer.

Can an 8GB Mac run Bolt E4B?

Not through Ollama. Every e4b tag is larger than the 5,600 MB budget of an 8GB Mac. A third-party 4-bit build fits: the lmstudio-community MLX 4-bit file is about 4.2 GB, and the unsloth UD-Q4_K_XL GGUF is about 4.8 GB. Expect text only and a short context window.

Did NVIDIA test Bolt E4B on Apple Silicon?

No. NVIDIA's model card lists Linux on Ampere, Hopper and Blackwell GPUs, and its published speed and memory tables cover GeForce RTX, RTX Spark and Jetson only. Apple Silicon support comes from Ollama, whose library page states E4B runs on Apple Silicon, and from the Gemma 4 architecture that llama.cpp and MLX already load.

Is the context window 64K or 128K?

NVIDIA says 64K, and its config and GGUF header both say 65,536 tokens. Ollama's release note and llama.cpp tags list 128K, its MLX tags list 64K, and the unsloth GGUF header says 131,072, inherited from Gemma 4. NVIDIA documents and benchmarks it at 64K, so longer contexts are untested.

What license does Bolt E4B use?

The OpenMDW License Agreement, version 1.1, which NVIDIA describes as allowing commercial and non-commercial use. The card adds that the underlying Google Gemma 4 model is governed by the Apache License, Version 2.0.

Sources

• NVIDIA-Nemotron-3.5-Bolt-E4B-4GB model card and Q4_K_M GGUF (NVIDIA, 2026)
• NVIDIA-Nemotron-3.5-Bolt-E4B-BF16 repository and config (NVIDIA, 2026)
• NVIDIA Nemotron 3.5 Bolt E4B 4GB research page (NVIDIA, October 23, 2026)
• Ollama library: nemotron-3.5-bolt and its tags (Ollama, 2026)
• Ollama v0.40.10 release notes (GitHub, October 23, 2026)
• unsloth NVIDIA-Nemotron-3.5-Bolt-E4B-GGUF (unsloth, 2026)
• lmstudio-community NVIDIA-Nemotron-3.5-Bolt-E4B-MLX-4bit (LM Studio community, 2026)
• llama.cpp build b11459 (GitHub, 2026)