Yes — with 92.5 GB to spare
Gemma 4 E2B at Q4_K_M fits your M4 Max · 128 GB entirely in unified memory at 8K context, at an estimated 106 tokens per second. There is room for its full 128K window.
Fully in unified memory
8K context
Q4_K_M · 2.9 GB
Apache 2.0
Released 2 Apr 2026
Vision
"E2B" is 2.3B effective, but the file holds 5B because of per-layer embeddings — size it as 5B. Text, image and audio in.
The VRAM budget
weights 2.9 GB
Weights 2.9 GB
KV cache @ 8K 0.07 GB
Runtime overhead 0.6 GB
Free 92.5 GB of 96.0 GB
Quantisation ladder
| Quant | Weights | Total @ 8K | Max context | Tok/s | Quality | Fit |
|---|---|---|---|---|---|---|
| F16 | 9.5 GB | 10.2 GB | 128K | 32 | Reference | Long context |
| Q8_0 | 5.0 GB | 5.7 GB | 128K | 60 | −0.1% ppl | Long context |
| Q6_K | 3.9 GB | 4.6 GB | 128K | 78 | −0.4% ppl | Long context |
| Q5_K_M | 3.4 GB | 4.0 GB | 128K | 91 | −0.8% ppl | Long context |
| Q4_K_M | 2.9 GB | 3.5 GB | 128K | 106 | −1.9% ppl | Recommended |
| Q3_K_M | 2.3 GB | 3.0 GB | 128K | 131 | −5.4% ppl | Long context |
| Q2_K | 2.0 GB | 2.7 GB | 128K | 153 | −15% ppl | Long context |
Quality is the published perplexity delta against f16 weights. Max context assumes an f16 KV cache; q8_0 roughly doubles it. This model interleaves sliding-window layers (512 tokens, 7 of 35 layers global), which is why its cache barely grows with context.
How to run it
$ pip install mlx-lm $ mlx_lm.generate --model mlx-community/gemma-4-E2B-it-4bit \ --max-tokens 512 --prompt "Hello"
Apple's own array framework. The fastest path on Apple Silicon. More on MLX.
01Download is 2.9 GB. Keep it on an SSD — a first load off a spinning disk takes minutes.
02macOS caps what the GPU may wire down at about 96.0 GB of 128 GB. Raising it with iogpu.wired_limit_mb is possible, and risky.
03There is room to go to the model's full 128K context on this card.