LFM3-1.7B-MLC

LiquidAI/LFM3-1.7B converted to MLC format so it runs in the browser with WebLLM on WebGPU. Weights, model libraries (.wasm) and the chat/tokenizer config are all in this repo, so WebLLM can load everything straight from here.

Built with mlc-llm for web-llm 0.2.85 (model library ABI v0_2_84).

Variants

quantization	folder	model library	needs	download
q6f16_a	LFM3-1.7B-q6f16_a-MLC/	libs/LFM3-1.7B-q6f16_a_cs1k-webgpu.wasm	WebGPU shader-f16	1458 MB

• q6f16_a — asym (K-quant-like) int6 everywhere

Use with WebLLM

import * as webllm from "@mlc-ai/web-llm";

const appConfig: webllm.AppConfig = {
  model_list: [
    {
      "model": "https://huggingface.co/CharlZKP/LFM3-1.7B-MLC/resolve/main/LFM3-1.7B-q6f16_a-MLC",
      "model_id": "LFM3-1.7B-q6f16_a-MLC",
      "model_lib": "https://huggingface.co/CharlZKP/LFM3-1.7B-MLC/resolve/main/libs/LFM3-1.7B-q6f16_a_cs1k-webgpu.wasm",
      "vram_required_MB": 3202.54,
      "low_resource_required": true,
      "overrides": {
        "context_window_size": 32768
      },
      "required_features": [
        "shader-f16"
      ]
    }
  ]
};

const engine = await webllm.CreateMLCEngine("LFM3-1.7B-q6f16_a-MLC", { appConfig });
const reply = await engine.chat.completions.create({
  messages: [{ role: "user", content: "Hello!" }],
});
console.log(reply.choices[0].message.content);

The same records are in webllm-model-records.json. model points at the folder of one quantization and model_lib at its .wasm; nothing else has to be hosted.

Files

LFM3-1.7B-q6f16_a-MLC/     weights (params_shard_*.bin), tensor-cache.json, mlc-chat-config.json, tokenizer
libs/                      model libraries (.wasm), one per quantization
webllm-model-records.json  ready-to-paste WebLLM ModelRecords

About the model

LiquidAI/LFM3-1.7B is a 1.73B-parameter foundational language model developed by Liquid AI featuring an innovative tri-hybrid architecture:

• 13 Gated Short-Convolution Layers ($k=7$) capturing local context with low compute and recurrent conv cache state.
• 5 Sliding-Window Attention (SWA) Layers ($W=256$) for efficient semi-local attention.
• 6 Full Global Attention Layers with zero-tail partial RoPE ($\theta=1,000,000$, partial rotary factor 0.5) supporting up to 256K native context.
• Sandwich RMSNorm (4 RMSNorms per layer) and head-wise gated attention suppressing activation outliers.

This model is also distributed in GGUF format as LiquidAI/LFM3-1.7B-GGUF.

This repository provides WebLLM / MLC weights and compiled WebGPU WGSL model libraries (.wasm), enabling real-time client-side browser execution via WebGPU with zero cloud dependencies.

---

Quantization Variants

q6f16_a (GGUF Q6_K Equivalent)

• Base Bit-width: 6-bit asymmetric quantization (6.625 bpw, ~1.43 GB).
• Format: 32 weights per sub-block (8-bit scale + 8-bit min), 256 weights per super-block (bf16 $d, dm$ scales).
• Precision Preservation: The 1D causal convolution weights (conv.conv.weight, shape $2048 \times 1 \times 7$) and attention head gating projections (self_attn.gate_proj, shape $16 \times 2048$) are strictly preserved in high-precision unquantized fp16 (no_quantization = True).
• WebGPU Buffer Compliance: Slices large embedding and vocabulary dequantization operations into 32,000-token chunks, strictly complying with the 128 MiB WebGPU maxStorageBufferBindingSize limit across desktop and mobile devices.

---

Use with WebLLM

import * as webllm from "@mlc-ai/web-llm";

const appConfig: webllm.AppConfig = {
  model_list: [
    {
      model: "https://huggingface.co/CharlZKP/LFM3-1.7B-MLC/resolve/main/LFM3-1.7B-q6f16_a-MLC",
      model_id: "LFM3-1.7B-q6f16_a-MLC",
      model_lib: "https://huggingface.co/CharlZKP/LFM3-1.7B-MLC/resolve/main/libs/LFM3-1.7B-q6f16_a_cs1k-webgpu.wasm",
      required_features: ["shader-f16"],
      vram_required_MB: 2150,
      low_resource_required: true,
      overrides: { context_window_size: 32768 }
    }
  ]
};

const engine = await webllm.CreateMLCEngine("LFM3-1.7B-q6f16_a-MLC", { appConfig });
const reply = await engine.chat.completions.create({
  messages: [{ role: "user", content: "Explain the benefits of hybrid convolution-attention models." }]
});
console.log(reply.choices[0].message.content);

---

Attribution & License

• Base Model: LiquidAI/LFM3-1.7B
• GGUF Reference: LiquidAI/LFM3-1.7B-GGUF
• Developer: Liquid AI
• License: LFM1.0 (license: other, license_name: lfm1.0)

How this was built

mlc_llm convert_weight / gen_config / compile (WebGPU) from a conversion pipeline with custom quantization presets, using an importance matrix from calibration text. The .wasm libraries contain the model's compiled kernels, including the custom dequantization kernel, and are checked against an official WebLLM library for ABI compatibility.

License

apache-2.0, inherited from LiquidAI/LFM3-1.7B.