kanha-liquid-lfm2.5-1.2b-instruct-sft-v1-MLC

Kanha Liquid LFM2.5 1.2B SFT, MLC WebGPU

Browser-ready MLC artifacts for Kanha's fine-tuned LiquidAI LFM2.5 1.2B model. The repository contains the complete q4f16_1 parameter cache, tokenizer, MLC chat configuration, and compiled WebGPU model library.

Immutable inputs

Runtime

• Quantization: q4f16_1
• Context window: 4,096 tokens
• Prefill chunk: 512 tokens
• State: hybrid attention KV cache and recurrent convolution state
• Model data: 663,326,625 bytes across 21 parameter shards
• WebGPU library: kanha-liquid-lfm2.5-1.2b-tuned-q4f16_1.wasm
• WebGPU library SHA-256: 9460955dcd10198c32b5d6946a68f59ada3443bc6400aa269bdd9ba7463f85c1
• Tested browser runtime: @mlc-ai/[email protected]

The model files are intentionally stored at the repository root. Pin an immutable 40-character repository revision in both URLs:

const model = "https://huggingface.co/Kanha-AI/kanha-liquid-lfm2.5-1.2b-instruct-sft-v1-MLC/resolve/<revision>/";
const modelLib = "https://huggingface.co/Kanha-AI/kanha-liquid-lfm2.5-1.2b-instruct-sft-v1-MLC/resolve/<revision>/kanha-liquid-lfm2.5-1.2b-tuned-q4f16_1.wasm";

Verification

The collected artifact was hash-checked against the conversion record before upload. A macOS WebGPU smoke test loaded it through a Web Worker with a 4,096-token context and produced a grounded answer with citations. That single smoke test confirms browser execution, not general model quality or behavior on other devices. See qualification.json and conversion-records/ for the recorded evidence.

License

This derivative is distributed under the included LFM Open License v1.0. Review LICENSE before use, especially its commercial-use conditions and attribution requirements.