CoolFace
Modelpublic

OsaurusAI/LFM2.5-230M-MXFP8

sourceHugging Faceotherupdated 1mo agoView on Hugging Face
2likes120downloads
Model Card

[image]

LFM2.5-230M · MXFP8

Official OsaurusAI MXFP8 build of LiquidAI/LFM2.5-230M (LFM Open License v1.0) — Liquid AI's 230M tiny hybrid model. Near-lossless 8-bit microscaled FP; runs on Apple Silicon via Osaurus / mlx_lm.

  • —~231 MB bundle (down from ~459 MB bf16) — small enough for the most constrained on-device use.
  • —MXFP8: microscaled FP8 (group-size 32) on the linear weights; short-conv kernels and norms kept fp16.
  • —Text-only, multilingual (en, ar, zh, fr, de, ja, ko, es).

Architecture

Familylfm2 (hybrid)
Layers14 — 8 short-conv (LIV) + 6 full-attention
Hidden1024 · vocab 65536 · tied embeddings
Cachehybrid (conv state + KV for attention layers)
ToolsLiquid Python-call format (lfm2 tool parser)

The short-conv (LIV) layers (conv.conv kernel + conv.in_proj/conv.out_proj) interleave with full-attention layers — verified coherent generation in mlx_lm after quantization.

Usage

bash
python -m mlx_lm generate --model OsaurusAI/LFM2.5-230M-MXFP8 --prompt "What is the capital of France?"

Or load in Osaurus for a local, no-setup agent loop.

Provenance

  • —Base: LiquidAI/LFM2.5-230M © Liquid AI — LFM Open License v1.0 (see LICENSE)
  • —Quantization: Osaurus · MXFP8 (microscaled FP8, group-size 32) · eric@osaurus.ai