[Feature]: Model Request: JetBrains Mellum2‑12B‑A2.5B‑Thinking & Mellum2‑12B‑A2.5B‑Instruct (MoE)

Suggestion Description

Model Request

Please add support for two JetBrains Mellum2 MoE models:

1. JetBrains/Mellum2‑12B‑A2.5B‑Thinking
2. JetBrains/Mellum2‑12B‑A2.5B‑Instruct

Model Info

• Total params: 12B, activated params: only 2.5B per token (MoE architecture)
• Context window: 131072 tokens
• License: Apache‑2.0
• HuggingFace: https://huggingface.co/JetBrains

Rationale

This MoE model delivers very strong code & reasoning performance. Compared to the existing GPT‑OSS 20B supported in FLM, Mellum2‑12B‑A2.5B runs much faster thanks to its low active parameter count, and would be a great fit for Ryzen AI XDNA2 NPU.

The Thinking variant brings advanced step‑by‑step reasoning capability, and the Instruct variant is good for general assistant usage. Both would be very valuable for local NPU‑powered development workflows.

I would love to run these two models via flm run on my Ryzen AI hardware.

Thanks for your great work on FastFlowLM!

Operating System

No response

GPU

No response

ROCm Component

No response

Start with the existing GPT-OSS support used by flm run and compare its model integration with the JetBrains Mellum2 HuggingFace entries. Verify both Thinking and Instruct variants, MoE behavior, the 131072-token context, and Ryzen AI NPU compatibility; done means both models run through flm run.