Yes — with 70.8 GB to spare
Gemma 3 1B at Q4_K_M fits your M2 Max · 96 GB entirely in unified memory at 8K context, at an estimated 398 tokens per second. There is room for its full 32K window.
Fully in unified memory
8K context
Q4_K_M · 0.6 GB
Gemma Terms of Use
Released Mar 2025
Text-only. Single KV head makes its cache almost free at long context.
The VRAM budget
weights 0.6 GB
Weights 0.6 GB
KV cache @ 8K 0.04 GB
Runtime overhead 0.6 GB
Free 70.8 GB of 72.0 GB
Quantisation ladder
| Quant | Weights | Total @ 8K | Max context | Tok/s | Quality | Fit |
|---|---|---|---|---|---|---|
| F16 | 1.9 GB | 2.5 GB | 32K | 120 | Reference | Long context |
| Q8_0 | 1.0 GB | 1.6 GB | 32K | 226 | −0.1% ppl | Long context |
| Q6_K | 0.8 GB | 1.4 GB | 32K | 293 | −0.4% ppl | Long context |
| Q5_K_M | 0.7 GB | 1.3 GB | 32K | 339 | −0.8% ppl | Long context |
| Q4_K_M | 0.6 GB | 1.2 GB | 32K | 398 | −1.9% ppl | Recommended |
| Q3_K_M | 0.5 GB | 1.1 GB | 32K | 491 | −5.4% ppl | Long context |
| Q2_K | 0.4 GB | 1.0 GB | 32K | 573 | −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, 1 global in 6), which is why its cache barely grows with context.
How to run it
$ pip install mlx-lm $ mlx_lm.generate --model mlx-community/gemma-3-1b-it-4bit \ --max-tokens 512 --prompt "Hello"
Apple's own array framework. The fastest path on Apple Silicon. More on MLX.
01Download is 0.6 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 72.0 GB of 96 GB. Raising it with iogpu.wired_limit_mb is possible, and risky.
03There is room to go to the model's full 32K context on this card.