baa-ai/Qwen3.6-27B-RAM-28GB-MLX

Qwen3.6-27B — 28GB (MLX)

Mixed-precision MLX build of Qwen/Qwen3.6-27B, prepared by baa.ai.

Built at the predicted global (quality max) operating point.

Metrics

Metric	Value
In-memory footprint	~28 GiB
Size on disk	27.8 GB
Average bits per weight	7.65
Group size	64
Framework	MLX (Apple Silicon)
Source	Qwen/Qwen3.6-27B (BF16, 55.6 GB)

Benchmarks

Reported under both decoding regimes following our internal evaluation policy that decoding choice can materially shift per-task accuracy on quantized checkpoints (see Notes).

Benchmark (greedy decode)	Score
WikiText-2 PPL (128×2048, seed 42)	4.856 ± 0.028
MMLU-Pro 140Q calibrated	54.3%
GPQA Diamond 198Q overall	37.9%
GPQA Diamond — Biology	31.6%
GPQA Diamond — Chemistry	40.9%
GPQA Diamond — Physics	36.0%

Benchmark (T=0.3, top_p=0.95, seed=123)	Score
GPQA Diamond 198Q overall	40.9%
GPQA Diamond — Biology	52.6%
GPQA Diamond — Chemistry	36.6%
GPQA Diamond — Physics	43.0%

Notes on decoding regime

This 28 GB build's ranking versus the companion 16 GB build flips between greedy and sampled decoding, particularly on Biology (greedy: this build loses by 16 pp; sampled: this build wins by 11 pp). We recommend evaluating against your specific deployment decoding configuration before committing to a variant. Best PPL and best MMLU-Pro under greedy belong to this build; under sampled decoding it is also the strongest on overall GPQA and the strongest on Biology.

Recommended inference settings

sampler_params = {
    "temperature": 1.0,
    "top_p": 0.95,
    "top_k": 40,
    "repetition_penalty": 1.1,
    "max_tokens": 8192,
}

Usage

from mlx_lm import load, generate
from mlx_lm.sample_utils import make_sampler, make_logits_processors

model, tokenizer = load("baa-ai/Qwen3.6-27B-RAM-28GB-MLX")

sampler = make_sampler(temp=1.0, top_p=0.95, top_k=40)
logits_processors = make_logits_processors(repetition_penalty=1.1)

prompt = tokenizer.apply_chat_template(
    [{"role": "user", "content": "Write a Python function that reverses a string."}],
    tokenize=False,
    add_generation_prompt=True,
)

response = generate(model, tokenizer, prompt=prompt, max_tokens=8192,
                    sampler=sampler, logits_processors=logits_processors)
print(response)

Hardware

• Apple Silicon Mac with ~36 GB unified memory recommended.

Variants

Variant	Size	Link
16 GB	18.2 GB	baa-ai/Qwen3.6-27B-RAM-16GB-MLX