Mellum2-12B-A2.5B-Thinking-mlx-4bit

This is a 4-bit MLX quantization of JetBrains/Mellum2-12B-A2.5B-Thinking, the reasoning-augmented Mixture-of-Experts coding assistant from JetBrains. It is derived from the full-precision jedisct1/Mellum2-12B-A2.5B-Thinking-mlx conversion.

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. It emits its reasoning inside <think>...</think> blocks before the final answer.

A 4-bit quant tuned for reliable tool calling

A naive flat 4-bit quantization of this model breaks tool calling. In testing, the model would emit edit-style tool calls whose JSON arguments contained mis-escaped quotes, so the server dropped them as malformed and agent runs stalled. The culprit was 4-bit precision on the attention and output layers, which is where the model copies literal source text verbatim into a tool argument.

This quantization avoids that. The 64 experts are the overwhelming majority of the weights, so they are quantized to 4 bits to keep the file small, while the parts that govern formatting and routing are kept at 8 bits:

• attention projections (q_proj, k_proj, v_proj, o_proj) — 8-bit
• token embeddings and lm_head — 8-bit
• the per-layer expert router (mlp.gate) — 8-bit
• the MoE experts (switch_mlp) — 4-bit

All quantization uses a group size of 64, which works out to about 4.85 bits per weight. The disk footprint is the same as a flat 4-bit model, but tool calling holds up.

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, completed multi-step coding tasks, 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).

If you want output that matches the original model as closely as possible and can spare the memory, use the 8-bit quantization instead.

Requirements

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:

pip install git+https://github.com/jedisct1/mlx-lm

Or run it directly with uv:

uvx --from git+https://github.com/jedisct1/mlx-lm mlx_lm.server

Use with mlx-lm

Quick test:

uvx --from git+https://github.com/jedisct1/mlx-lm \
  mlx_lm.generate --model jedisct1/Mellum2-12B-A2.5B-Thinking-mlx-4bit \
  --prompt "How are you doing today?" \
  --max-tokens 16384 \
  --temp 0.6 --top-p 0.95 --top-k 20

Starting the server:

uvx --from git+https://github.com/jedisct1/mlx-lm \
  mlx_lm.server --model jedisct1/Mellum2-12B-A2.5B-Thinking-mlx-4bit \
  --max-tokens 16384 \
  --temp 0.6 --top-p 0.95 --top-k 20

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

Using this setup with the Swival.dev harness

Install swival.dev:

uv tool install swival

Then point it at the running server:

swival --provider llamacpp --model jedisct1/Mellum2-12B-A2.5B-Thinking-mlx-4bit

License

Apache 2.0, inherited from the original model.