mlx-community/Mellum2-12B-A2.5B-Thinking-8bit

Mellum2 12B A2.5B Thinking - 8-bit MLX

This is an 8-bit affine MLX quantization of JetBrains/Mellum2-12B-A2.5B-Thinking.

Mellum2 Thinking is a reasoning-augmented Mixture-of-Experts assistant model. It has 64 experts with 8 active per token, a 131,072-token context window, and emits reasoning in <think>...</think> blocks before the final answer.

Conversion details

• Source: JetBrains/Mellum2-12B-A2.5B-Thinking
• Format: MLX safetensors
• Quantization: affine, 8 bits, group size 64
• License: Apache-2.0
• EOS token: <|im_end|> (token ID 28)

The upstream config.json and generation_config.json identify token ID 0 as the EOS token, while the tokenizer identifies <|im_end|> (ID 28) as EOS. This conversion uses token ID 28 so MLX generation stops at the end of the assistant turn.

Usage

pip install -U mlx-lm

mlx_lm.chat \
  --model mlx-community/Mellum2-12B-A2.5B-Thinking-8bit \
  --max-tokens 8192 \
  --temp 0.6 \
  --top-p 0.95

Benchmarks

Benchmarked with oMLX 0.5.7 and MLX-LM 0.31.3 on an Apple M5 Max MacBook Pro with 128 GB unified memory, using the built-in Python-code workload and 128 generated tokens. Results vary with hardware, runtime versions, context length, and generation settings.

Prompt tokens	TTFT	Prefill	Generation	End-to-end	Peak memory
4,096	926.4 ms	4,421.3 tok/s	132.3 tok/s	1.9 s	12.7 GB
16,384	3,205.2 ms	5,111.8 tok/s	121.8 tok/s	4.3 s	13.0 GB
32,768	8,171.5 ms	4,010.0 tok/s	112.9 tok/s	9.3 s	13.4 GB
65,536	22,286.9 ms	2,940.6 tok/s	92.9 tok/s	23.7 s	14.1 GB
131,072	67,675.6 ms	1,936.8 tok/s	56.0 tok/s	70.0 s	15.5 GB

At a 4,096-token prompt, batch generation achieved 176.4, 189.3, and 222.8 aggregate generation tok/s at 2, 4, and 8 concurrent requests, respectively, compared with 132.3 tok/s for one request.

Model provenance

For the original model card, training details, benchmark results, and usage guidance, see the upstream JetBrains checkpoint.