This is an MLX version of JetBrains/Mellum2-12B-A2.5B-Instruct, the instruction-tuned Mixture-of-Experts coding assistant from JetBrains. The weights are kept in their native bfloat16 precision, so the model behaves exactly like the original checkpoint.

Unlike its sibling Mellum2-12B-A2.5B-Thinking, the Instruct model answers directly without emitting a <think> reasoning block, which makes it faster and lighter on tokens for straightforward coding and tool-use tasks.

Mellum 2 uses 64 experts with 8 active per token (about 2.5B active parameters out of 12B), a mix of sliding-window and full-attention layers, and a 131,072-token context window.

Tool calling was verified end to end against a live mlx_lm.server driven by the swival agent harness: across repeated runs the model issued well-formed read_file, edit_file, write_file, list_files, and shell-command calls and never produced a malformed tool call. Generation stops cleanly on <|im_end|> (the eos_token_id is set to [0, 28], which is what lets agent harnesses see a proper tool_calls finish reason — the upstream checkpoint ships eos_token_id: 0, which never fires on a chat turn and leaves tool calls running past the token limit).

If you want the same model with a smaller footprint:

The mellum architecture is not supported by the stock mlx-lm code yet.

Until it is supported upstream, install this fork of mlx-lm from source:

Or run it directly with uv:

Quick test:

Starting the server:

The recommended sampling settings from JetBrains are temperature=0.6, top_p=0.95, top_k=20.

Install swival.dev:

Then point it at the running server:

Apache 2.0, inherited from the original model.