LFM2.5-8B-A1B 8-bit XL MLX
LFM2.5-8B-A1B by Liquid AI — a hybrid conv + attention MoE built for on-device deployment (extended pre-training + RL on the LFM2 architecture), quantized for Apple Silicon with the BaseQuant_XL 8-bit recipe.
8.3B total / 1.5B active parameters. Designed to power on-device personal assistants: chained tool calls, complex instruction following, agentic tasks — at a memory footprint that fits 16 GB Macs.
About XL Quantization
BaseQuant_XL is a fully data-agnostic, static quantization. No calibration dataset, no sensitivity analysis, no importance matrix. Precision is allocated purely by architectural role — routing-critical layers get higher precision, bulk expert parameters get lower precision. The result is a transparent, faithful capture of the source model.
Data-dependent calibration quantizations (iMatrix, AWQ, GPTQ, oQ, oQ4e, etc.) use a calibration set to guide bit allocation. This can produce a skewed representation of the model: domains well-represented in the calibration data (English, popular topics, public or leaked benchmarks) are preserved better, while underrepresented domains (non-English languages, niche use cases, your own data) are preserved worse. XL avoids this trade-off entirely — it generalizes honestly because it is never fit to any particular data distribution.
Quickstart
pip install -U mlx-lm
mlx_lm.chat --model leonsarmiento/LFM2.5-8B-A1B-8bit-XL-mlx
Thinking is native to LFM2.5 — reasoning traces are emitted inline in the response.
Quantization Strategy
| Layers | Bits | Rationale |
|---|---|---|
feed_forward.gate (router) |
bf16 | Routing decisions — any quantization noise here causes expert misrouting |
embed_tokens (tied = output head), attention, short-conv, dense MLP, routed experts |
8-bit | Near-lossless across every remaining layer — max-fidelity XL |
- Bits per weight: 8.501 · Total size: ~9.0 GB (2 shards) · Group size: 64
Note: tie_word_embeddings=true — the 8-bit embedding doubles as the output head (same treatment as our Gemma 4 XL builds).
Liquid AI also publishes their own uniform 8-bit MLX build (LFM2.5-8B-A1B-MLX-8bit); this XL variant differs by keeping the router gate in bf16.
Recommended Inference Parameters
Per Liquid AI's model card / shipped generation_config.json:
| Parameter | Value |
|---|---|
temperature |
0.2 |
top_k |
80 |
repetition_penalty |
1.05 |
Model Overview
| Property | Value |
|---|---|
| Architecture | Lfm2MoeForCausalLM (lfm2_moe) |
| Parameters | 8.3B total / 1.5B active |
| Layers | 24 — 18 double-gated short conv + 6 GQA full attention |
| Experts | 32 routed, 4 active/token + expert bias |
| Training budget | 38 trillion tokens |
| Context window | 128,000 tokens |
| Vocabulary | 128,000 |
| Languages | en, ar, zh, fr, de, ja, ko, es, pt, it |
| License | LFM Open License (lfm1.0) |
Source Model Benchmarks (from Liquid AI)
| Benchmark | LFM2-8B-A1B | LFM2.5-8B-A1B | Δ |
|---|---|---|---|
| AA-Omniscience Index | -78.42 | -24.70 | +53.62 |
| IFEval | 79.44 | 91.84 | +12.40 |
| IFBench | 26.00 | 56.47 | +30.47 |
| Multi-IF | 58.54 | 79.93 | +21.39 |
| MATH500 | 74.80 | 88.76 | +13.96 |
| AIME25 | 20.00 | 42.53 | +22.53 |
| BFCLv3 | 45.07 | 64.36 | +19.29 |
| BFCLv4 | 25.52 | 48.50 | +22.98 |
| Tau² Telecom | 13.60 | 88.07 | +74.47 |
| Tau² Retail | 7.02 | 39.82 | +32.80 |