baa-ai/Qwen3.6-27B-RAM-16GB-MLX

Qwen3.6-27B — 16GB (MLX)

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

Built at the predicted local (capability) operating point.

Metrics

Metric	Value
In-memory footprint	~16 GiB
Size on disk	18.2 GB
Average bits per weight	4.23
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)	5.221 ± 0.031
MMLU-Pro 140Q calibrated	52.1%
GPQA Diamond 198Q overall	37.9%
GPQA Diamond — Biology	47.4%
GPQA Diamond — Chemistry	35.5%
GPQA Diamond — Physics	38.4%

Benchmark (T=0.3, top_p=0.95, seed=123)	Score
GPQA Diamond 198Q overall	43.4%
GPQA Diamond — Biology	42.1%
GPQA Diamond — Chemistry	33.3%
GPQA Diamond — Physics	54.7%

Notes on decoding regime

Sampled decoding adds ~5 pp to overall GPQA accuracy on this checkpoint and reorders per-subject performance versus greedy. Specifically, the per-subject ranking between this 16 GB build and the companion 28 GB build inverts under T=0.3 sampling on Biology (greedy: 16 GB wins by 16 pp; sampled: 28 GB wins by 11 pp). Both checkpoints are public for two reasons: (a) the inversion is an interesting empirical demonstration that quantization × decoding produce coupled effects we are still characterising, and (b) Physics accuracy on this 16 GB build under sampled decoding (54.7%) exceeds the 28 GB build under either regime, so this checkpoint genuinely wins on at least one task in at least one regime. We recommend evaluating against your specific deployment decoding configuration before committing to a variant.

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-16GB-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 ~21 GB unified memory recommended.

Variants

Variant	Size	Link
28 GB	27.8 GB	baa-ai/Qwen3.6-27B-RAM-28GB-MLX