The LLMCheck index estimates Nemotron 3.5 Bolt E4B at ~114 tok/s on an M3 Max at Q4_K_M, a figure from its published bandwidth model, not a measured run. It needs about 6 GB of unified memory. It is NVIDIA's 4B dense model, released under the OpenMDW licence.
Want to run Nemotron 3.5 Bolt E4B faster — or step up to bigger models? Find your Mac →This is an LLMCheck index estimate from its published two-term bandwidth model, not a measured run; the index labels it so everywhere it appears. Own this Mac? Submit a real run →
| Metric | Value |
|---|---|
| Tokens per second | estimated~114 tok/s |
| Provenance | estimate · how estimates are made |
| Quantization | Q4_K_M |
| Minimum RAM | 6 GB |
| Parameters | 4B |
| Chip bandwidth | 400 GB/s |
Each row as the index holds it, with its source. “Mac RAM” is the memory of the Mac a run used, not the model’s requirement (6 GB). The headline figure follows the reference rule: the lower-median measured plain-decode run at 5 bits or fewer (a run whose quant was not stated counts only when no stated-quant run qualifies), else the estimate; a run marked “not used to anchor estimates” is listed here but never becomes the headline.
| Speed | Quant | Engine | Mac RAM | Context | Source | Note |
|---|---|---|---|---|---|---|
| community72.98 tok/s | 4-bit | MLX | 36 GB | short chat prompt, 256 generated, greedy | huggingface.co | The 36 GB M3 Max is the binned 30-core-GPU, 300 GB/s part. mlx-vlm 0.7.10 on MLX 0.32.4, 4-bit affine g64; decode-only gen_tps_reported from results/kl_results_release.json (the card table's 68.1 is wall-clock with prefill). The quantizer reports only 71.7% top-1 agreement with BF16 for this build.; binned GPU |
| community55.22 tok/s | 6-bit | MLX | 36 GB | short chat prompt, 256 generated, greedy | huggingface.co | The 36 GB M3 Max is the binned 30-core-GPU, 300 GB/s part. mlx-vlm 0.7.10 on MLX 0.32.4, 6-bit affine g64; decode-only gen_tps_reported (wall-clock 46.8). 92.5% top-1 agreement with BF16, the quality-safe build.; binned GPU |
| community45.26 tok/s | 8-bit | MLX | 36 GB | short chat prompt, 256 generated, greedy | huggingface.co | The 36 GB M3 Max is the binned 30-core-GPU, 300 GB/s part. mlx-vlm 0.7.10 on MLX 0.32.4, 8-bit affine g64; decode-only gen_tps_reported (wall-clock 36.6).; binned GPU |
| community26.53 tok/s | BF16 | MLX | 36 GB | short chat prompt, 256 generated, greedy | huggingface.co | The 36 GB M3 Max is the binned 30-core-GPU, 300 GB/s part. mlx-vlm 0.7.10 on MLX 0.32.4; decode-only gen_tps_reported (wall-clock 11.4).; binned GPU |
Ollama — install it (brew install ollama or the app from ollama.com), then pull the model by its tag, checked on ollama.com on 2026-10-24:
MLX — pick a build from Hugging Face’s MLX listings, then run it with its repo id:
Downloads: Ollama page → · model card →
Prefer a GUI? LM Studio downloads GGUF and MLX builds in one click. Step by step: install Ollama on a Mac.
NVIDIA's OpenMDW-licensed edge model (23 Oct 2026), RL post-trained from Gemma 4 E4B for multi-turn tool use and instruction following: 8B weights, 4B effective, 64K context per NVIDIA. '4GB' is NVIDIA's VRAM target; the official Q4_K_M file is 6.18 GB, and 3–4-bit builds of ~4.2 GB fit an 8 GB Mac. Ollama runs it on Apple Silicon as nemotron-3.5-bolt:e4b (the untagged name is the 74B model). A community MLX run measured 55 tok/s at 6-bit on a 36 GB M3 Max (the 4-bit build reached 73, but its quantizer reports only 71.7% top-1 agreement with BF16), and a community oMLX run measured 46 at 8-bit on an M5 Pro.
OpenMDW licence · 4B · 64K context · LLMCheck capability 17/50 · see its leaderboard rank →
Nemotron 3.5 Bolt E4B needs about 6 GB of unified memory. These current Apple Silicon Macs have the headroom to run it comfortably:
As an Amazon Associate, LLMCheck earns from qualifying purchases. These affiliate links cost you nothing extra and help keep the LLMCheck index free.
See how Nemotron 3.5 Bolt E4B stacks up against other models on your specific Mac hardware.
Open Compare Tool Full Leaderboard