CoolFace
Modelpublic

JANGQ-AI/LFM2.5-8B-A1B-JANG_2L

sourceHugging Faceotherupdated 19d agoView on Hugging Face
0likes194downloads
Model Card

<p align="center"> <a href="https://vmlx.net"><img src="vmlx-app.png" alt="vMLX — run JANG models on Apple Silicon" width="820"></a> </p>

<h3 align="center">⚡ All JANG models are meant to be run in <a href="https://vmlx.net">vMLX</a></h3>

LFM2.5-8B-A1B-JANG_2L

JANG_2L conversion of LiquidAI/LFM2.5-8B-A1B, built for Apple Silicon inference through JANG-aware MLX/vMLX runtimes.

This bundle is not a plain MLX 2-bit quant. It uses JANG importance allocation over MLX affine quantized tensors, with higher precision reserved for runtime-sensitive tensors.

Format

  • —Format: JANG affine
  • —Profile: JANG_2L
  • —Quantization backend: mx.quantize
  • —Group size: 64
  • —Actual bits from jang_config.json: 2.37
  • —Bit widths used: 2, 6, 8
  • —Passthrough bit width: 16
  • —Local size before upload: 2.9G
  • —JANG runtime weight size metadata: 2.84 GB
  • —Source model: LiquidAI/LFM2.5-8B-A1B

Runtime capability stamp:

json
{
  "reasoning_parser": "qwen3",
  "tool_parser": "lfm2",
  "think_in_template": false,
  "supports_tools": true,
  "supports_thinking": true,
  "family": "lfm2_moe",
  "modality": "text",
  "cache_type": "hybrid"
}

Runtime

Use a JANG-aware MLX/vMLX runtime. The model has hybrid cache behavior: attention layers use KV cache, while LIV convolution layers use convolution/state cache.

Example with the local JANG tools runtime:

sh
python -m jang_tools inference \
  --model OsaurusAI/LFM2.5-8B-A1B-JANG_2L \
  --prompt "What is 2+2? Answer briefly." \
  --max-tokens 128 \
  --temperature 0

Chat Template And Reasoning

The bundled chat_template.jinja uses Liquid's ChatML-like format:

  • —User and assistant turns use <|im_start|> / <|im_end|>.
  • —The generation prompt ends at <|im_start|>assistant\n; it does not pre-open <think>.
  • —Assistant reasoning may appear inside <think>...</think>.
  • —Tool calls use Liquid's Python-call list format inside <|tool_call_start|> and <|tool_call_end|>.

For this bundle, think_in_template=false is intentional. Runtime code should parse reasoning if the model emits it, but should not force a second reasoning prefix.

Verification

Local smoke run on the converted bundle:

  • —Prompt: What is 2+2? Answer briefly.
  • —Result: output closed <think>...</think> and answered 2 + 2 = 4.
  • —Reported generation speed: 206.886 tok/s
  • —Load time: 1.946 s
  • —Peak RSS: 3887 MB

This is a smoke test, not a benchmark suite or accuracy evaluation.

Korean

이 모델은 LiquidAI/LFM2.5-8B-A1B를 JANG2L 형식으로 변환한 Apple Silicon용 번들입니다. `thinkin_template=false가 의도된 설정이며, 런타임은 모델이 생성한 <think>...</think>`를 파싱하되 별도의 reasoning 접두어를 강제로 추가하지 않아야 합니다.