MiniCPM5-1B-Agentic-Tooluse-v3-q4f16_1-MLC

MiniCPM5-1B-Agentic-Tooluse-v3 — MLC / WebLLM

MLC q4f16_1 conversion of the author's merged FP16 v3 checkpoint, the full-precision sibling of the requested GGUF model. Converted directly from FP16, without dequantizing a GGUF file.

This repository includes the quantized weight shards, tokenizer, chat configuration, and a model-specific WebGPU WebAssembly library.

Use in WebLLM

npm install @mlc-ai/[email protected]

Use the included webllm-example.js in a browser application served over HTTPS or localhost. It exports appConfig, formatToolSystem, and an example() function that loads the model and asks for the weather in Paris.

import { example } from './webllm-example.js';
const { engine, response, xml } = await example();
console.log(xml);

Tool schemas are inserted in the system prompt using the checkpoint's XML tool format. Pass extra_body: { enable_thinking: false } on each request. Output is returned as message text, for example:

<function name="get_weather"><param name="city">Paris</param></function>

This is not an automatic adapter for the OpenAI tools / tool_calls response fields. The example extracts the first completed XML function block. The upstream model can repeat or continue after a call, so use that completed block as the boundary and validate its name and arguments before dispatch. No tool is executed by the example.

The original checkpoint advertises a longer context, but this artifact is configured and compiled for a 4,096-token browser context. Quantization can change outputs; no upstream benchmark score is claimed for this conversion.

Validation

Tested in Chrome 153 with WebLLM 0.2.85 on an NVIDIA RTX 3090. The model loaded all 15 shards, compiled its GPU shaders, generated the get_weather call for Paris shown above, and stopped with finish_reason: "stop". All shard sizes, MD5 checksums, and 245 tensor bounds were verified. This is a functional smoke test, not a tool-calling accuracy benchmark.

Provenance and reproduction

• Source revision: 6ec64641849bea2b3b599641b6a2d7155ff09039.
• Conversion: official MLC convert_weight, CPU, q4f16_1.
• Build: local Ubuntu 24.04 under WSL; no paid cloud job used.
• MLC package: mlc-llm-nightly-cpu==0.26.dev6.
• TVM package: mlc-ai-nightly-cpu==0.26.dev246.
• TVM FFI: apache-tvm-ffi==0.1.13.
• Runtime sources: MLC 9fa644f54b04983adea4d0168f49fc6af4a893ba, TVM 837cb9de1127b48ce48e4cefe09e83215b9d4ba7.
• Three MLC compiler passes restored from upstream 2f78caa4db0f90730a11ee3bb5cbd5f23bf67f9f to match the TVM wheel's buffer-map API; see build/pin_compiler_passes.py.
• Emscripten: 6.0.9.
• WebAssembly runtime build includes upstream custom_allocator.cc, omitted by its amalgamation, and links through em++ for C++ runtime support.

The build/ directory contains the reproduction scripts and exact Python package versions. See validation.json for the checks actually performed on this artifact.

Attribution

Original fine-tuning: ewin-reg. Base model: OpenBMB MiniCPM5-1B. Conversion and WebLLM packaging: marioysc. Apache-2.0; see LICENSE.