sandseppel/lfm3-wsp-tactical-executorch

LFM3-1.7B WSP · Tactical ExecuTorch (Android)

Weight-Space-Programmed (WSP) variant of LiquidAI/LFM3-1.7B, exported for on-device inference with ExecuTorch + XNNPACK and react-native-executorch.

The world-knowledge neurons in the SwiGLU MLPs (layers 8–21) were pruned and the model was tactically quantized per-layer (4/8-bit mixed). Reasoning and tool-calling ability are intentionally preserved; world knowledge is intentionally sacrificed.

Files

1_7b/
├── xnnpack/
│   └── lfm3_wsp_tactical_xnnpack.pte   # 928.99 MB, Tier-2 G64, XNNPACK 4-bit
├── tokenizer.json
└── tokenizer_config.json               # includes embedded chat_template

Model details

Property	Value
Base model	LiquidAI/LFM3-1.7B
Active parameters	1,434,675,968 (82.99 % of baseline)
World knowledge	intentionally removed (−91.7 % on knowledge probe)
Reasoning	89.58 % (baseline 93.75 %)
Tool calling	100 % (baseline 95 %)
PTE size	928.99 MB
Context length	1024 tokens
Quantization	8-bit activations, mixed 4/8-bit weights, Tier-2 group size 64
Attention backend	llama::update_cache, llama::custom_sdpa, llama::sdpa_with_kv_cache
Runtime	ExecuTorch 1.x portable + XNNPACK, Android arm64
Tokenizer	128k vocab, chat template included

Phone benchmark

Samsung Galaxy S20 FE (SM-G781U1), react-native-executorch, 128 new tokens, identical prompt:

Model	Tok/s
Official LFM3-1.7B 8da4w baseline	16.8 / 16.0 / 16.6 / 15.2
WSP Tactical G64 (this repo)	18.0 / 17.9 / 17.4 / 16.7

Usage with react-native-executorch

import { useLLMChatSession } from 'react-native-executorch';

const BASE = 'https://huggingface.co/sandseppel/lfm3-wsp-tactical-executorch/resolve/main';
const session = useLLMChatSession({
  modelPath: `${BASE}/1_7b/xnnpack/lfm3_wsp_tactical_xnnpack.pte`,
  tokenizerPath: `${BASE}/1_7b/tokenizer.json`,
  tokenizerConfigPath: `${BASE}/1_7b/tokenizer_config.json`,
});

Make sure the ExecuTorch native runtime in your app is built with:

• the LLM custom ops (executorch.extension.llm.custom_ops)
• the quantized kernels
• XNNPACK backend

Notes

• This is a research / hobby artifact, not an official Liquid AI release.
• The model deliberately does not answer general knowledge questions reliably. It is optimized for reasoning, instruction following and tool calls.
• The PTE is exported for a static max context of 1024 tokens.
• Exported with ExecuTorch from the calibrated WSP checkpoint. Validation: prefill/decode logits cosine 0.961/0.963 vs. the exact G32 calibration, identical argmax, 34-token identical greedy prefix.

License

Inherits the license and usage restrictions of LiquidAI/LFM3-1.7B.