Q6 is the recommended quant for this model on Apple Silicon. It runs comfortably within 128 GB, delivers near-lossless quality vs Q8, and achieves significantly better throughput due to lower memory pressure.

Benchmark (Apple M4 Max, 128 GB)

Metric	Q6	Q8
Prompt processing	43.4 tok/s	0.9 tok/s
Generation	54.7 tok/s	32.2 tok/s
Peak memory	~43 GB	~106 GB

Q6 is 48x faster on prompt processing and nearly 2x faster on generation compared to Q8 on this machine, due to the large difference in memory pressure.

About the model

This model is based on Qwen3-30B-A3B-Thinking-2507 (MoE, 128 experts, 10 activated) with the Brainstorm 40x adapter by DavidAU, which expands and recalibrates the reasoning centers of the model to improve depth, coherence, and detail in outputs.

Key characteristics:

53B total parameters, ~3.3B activated per token (MoE architecture)
256k native context window
Thinking mode only — outputs include a <think> reasoning trace before the final response
Excels at coding and programming across all major languages
Strong performance on reasoning, math, and agentic tasks

For full model details, settings, and recommended samplers, see the original model card.

Usage

pip install mlx-lm

bash

mlx_lm.generate \
  --model TxemAI/Qwen3-53B-A3B-2507-TOTAL-RECALL-v2-MLX-6bit \
  --prompt "Write a Python function to sort a list of dicts by a key" \
  --max-tokens 4096

python

from mlx_lm import load, generate

model, tokenizer = load("TxemAI/Qwen3-53B-A3B-2507-TOTAL-RECALL-v2-MLX-6bit")

prompt = "Write a Python function to sort a list of dicts by a key"
messages = [{"role": "user", "content": prompt}]
text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)

response = generate(model, tokenizer, prompt=text, max_tokens=4096, verbose=True)

Recommended inference settings

Parameter	Value
Temperature	0.3 – 0.6
Top-K	20 – 40
Top-P	0.95
Min-P	0.05
Rep penalty	1.05 – 1.1
Max tokens	8k – 32k (thinking model, needs space)

For complex coding or reasoning tasks, use higher token limits (32k–81k).

Other formats

Format	Link
Original (safetensors, full precision)	DavidAU/Qwen3-53B-A3B-2507-THINKING-TOTAL-RECALL-v2-MASTER-CODER
GGUF	mradermacher/Qwen3-53B-A3B-2507-TOTAL-RECALL-v2-MASTER-CODER-GGUF
MLX Q8 (this collection)	TxemAI/Qwen3-53B-A3B-2507-TOTAL-RECALL-v2-MLX-8bit

TxemAI — MLX models for Apple Silicon

This model is part of the TxemAI collection of MLX-quantized models optimized for local inference on Apple Silicon, with a focus on data sovereignty and enterprise use cases.