Is MiniCPM-V 4.7 supported?

[TomLucidor]

They added a new tooling to go with their MTP, which apparently speeds things up by like around 30-50%? https://github.com/tc-mb/llama.cpp-omni/blob/master/docs/les/lookahead-expert-scheduling.md

[raullenchai]

Next release!

[TomLucidor]

@raullenchai also please include this one (quantized version) https://huggingface.co/LiquidAI/LFM2.5-24B-A2B

And as for the GPU side this looks interesting simplexai-labs/PolyStrata#1 (comment) simplexai-labs/PolyStrata#4 (comment)

[raullenchai]

@TomLucidor LFM2.5-24B-A2B now has a first-class alias on main (#4852), using the official 4-bit MLX build LiquidAI/LFM2.5-24B-A2B-MLX-4bit with tool calling and reasoning parsing preconfigured:

rapid-mlx serve lfm2.5-24b-a2b-4bit

It will ship in the next release.

[raullenchai]

Yes. MiniCPM-V 4.7 support was merged to main in #4745. It will ship in the next release after v0.16.3. There are two aliases, and both are marked experimental:

• minicpm-v4.7-1b-8bit (mlx-community/MiniCPM-V-4.7-1B-8bit)
• minicpm-v4.7-35b-a3b-mxfp4 (mlx-community/MiniCPM-V-4.7-35B-A3B-mxfp4)

Both accept image input and use structured tool calls:

rapid-mlx serve minicpm-v4.7-1b-8bit

The MTP and lookahead expert-scheduling speedups from the link are not enabled for these aliases yet. They are served with plain decode.

For LFM2.5-24B-A2B, please open a model-support issue so we can track it there.