mlx-community/LFM2.5-24B-A2B-OptiQ-4bit
Built with mlx-optiq, the MLX-native toolkit to quantize, fine-tune, and serve LLMs locally on Apple Silicon, no PyTorch and no cloud. All OptiQ quants 路 Docs 路 LFM2.5 family
An OptiQ mixed-precision quant of LiquidAI/LFM2.5-24B-A2B, the largest model in the LFM2.5 family. It is 16.7 GB on disk, down from 47.9 GB at bf16, with a 128k context.
LFM2.5-24B-A2B is a sparse mixture of experts on Liquid's hybrid backbone. Convolutional blocks alternate with full attention, so only 10 of its 40 layers keep a KV cache. Each token is routed to 4 of 64 experts. OptiQ measures each layer's sensitivity and assigns its bit-width from that.
What it is
| Property | Value |
|---|---|
| Base | LiquidAI/LFM2.5-24B-A2B |
| Architecture | lfm2_moe: hybrid conv + attention, 40 layers (10 attention), 64 experts, 4 active |
| Method | OptiQ mixed-precision, sensitivity-driven (uniform-4-bit reference) |
| On disk | 16.7 GB (bf16: 47.9 GB) |
| Context | 128k |
Most of the routed expert weights sit at 4-bit: 87 of the 114 fused expert projections. Attention, the routers and nearly all of the convolution blocks stay at 8-bit.
Capability Score
Six-metric mean, the standard OptiQ eval, run in reasoning mode because the model always thinks before it answers. The baseline is a uniform 4-bit quant of the same model, scored the same way.
| Metric | OptiQ | Uniform 4-bit |
|---|---|---|
| MMLU (generative) | 81.1% | 80.3% |
| GSM8K | 94.4% | 93.2% |
| IFEval (strict) | 88.0% | 84.5% |
| BFCL-V3 simple (200 calls) | 91.5% | 89.5% |
| HumanEval (164 problems, pass@1) | 91.5% | 67.1% |
| HashHop (long-context retrieval) | 1.0% | 0.0% |
| Capability Score (mean of 6) | 74.59 | 69.10 |
The OptiQ quant is ahead on every metric. The big gap is HumanEval: the uniform quant often writes a correct function and then keeps re-checking it until it runs out of tokens, even with a 16k reasoning budget. A reply that never finishes counts as a failure. HashHop is the weak spot for both, because the model copies the lookup table into its reasoning and runs out of budget before it answers.
Run it
pip install mlx-optiq
optiq serve --model mlx-community/LFM2.5-24B-A2B-OptiQ-4bit
That gives you an OpenAI and Anthropic compatible endpoint with mixed-precision KV cache (kv_config.json in this repo), tool calling, and prompt caching.
The model reasons before every answer, so give it a generous max_tokens. Liquid's recommended sampling (temperature 0.1, top_k 50, repetition_penalty 1.05) ships in generation_config.json, and optiq serve applies it automatically.
Or from Python:
from mlx_lm import load, generate
model, tokenizer = load("mlx-community/LFM2.5-24B-A2B-OptiQ-4bit")
prompt = tokenizer.apply_chat_template(
[{"role": "user", "content": "Summarise this contract clause in two sentences."}],
add_generation_prompt=True, tokenize=False,
)
print(generate(model, tokenizer, prompt=prompt, max_tokens=4096))
Links
- Project website: mlx-optiq.com
- All OptiQ quants: mlx-optiq.com/models
- PyPI: pypi.org/project/mlx-optiq
- Base model: LiquidAI/LFM2.5-24B-A2B
- Downloads last month
- 607
4-bit