Yes — with 381.8 GB to spare
LFM2.5 2.6B at Q4_K_M fits your M3 Ultra · 512 GB entirely in unified memory at 8K context, at an estimated 294 tokens per second. There is room for its full 128K window.
Fully in unified memory
8K context
Q4_K_M · 1.5 GB
LFM Open License v1.0
Released 28 Jul 2026
New this month
Not in the Ollama library
Convolution-heavy hybrid for CPUs and NPUs: 22 of 30 blocks keep no KV cache at all.
The VRAM budget
weights 1.5 GB
Weights 1.5 GB
KV cache @ 8K 0.13 GB
Runtime overhead 0.6 GB
Free 381.8 GB of 384.0 GB
Quantisation ladder
| Quant | Weights | Total @ 8K | Max context | Tok/s | Quality | Fit |
|---|---|---|---|---|---|---|
| F16 | 5.0 GB | 5.8 GB | 128K | 89 | Reference | Long context |
| Q8_0 | 2.7 GB | 3.4 GB | 128K | 167 | −0.1% ppl | Long context |
| Q6_K | 2.1 GB | 2.8 GB | 128K | 217 | −0.4% ppl | Long context |
| Q5_K_M | 1.8 GB | 2.5 GB | 128K | 251 | −0.8% ppl | Long context |
| Q4_K_M | 1.5 GB | 2.2 GB | 128K | 294 | −1.9% ppl | Recommended |
| Q3_K_M | 1.2 GB | 2.0 GB | 128K | 364 | −5.4% ppl | Long context |
| Q2_K | 1.1 GB | 1.8 GB | 128K | 425 | −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. Only 8 of its 30 blocks keep a per-token KV cache; the rest are linear-attention, Mamba or convolution blocks with a fixed-size state.
How to run it
$ pip install mlx-lm $ mlx_lm.generate --model mlx-community/LFM2.5-2.6B-4bit \ --max-tokens 512 --prompt "Hello"
Apple's own array framework. The fastest path on Apple Silicon. More on MLX.
01Download is 1.5 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 384.0 GB of 512 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.