Qwen performance on M1 Max 64GB Macbook

qwen_m1_macbook.md

Prompt: "Please review the home.nix configuration and note anything that can be done to simplify the nix configuration"

All the prompts review 2-3 .nix files and recommend 6-10 improvements on them.

The following are raw notes about how each model performs.

Hardware: M1 Max (10 CPU) 64GB RAM Macbook (Tahoe 26.5)

Notes:

• GGUF models are portable and can run on various platforms, while MLX models are optimized for Apple Silicon. https://atomic.chat/blog/guides/gguf-vs-mlx

Best model is also the slowest:

5tok/s

llama-server -hf unsloth/Qwen3.6-27B-MTP-GGUF:Q8_0 \
	--spec-type draft-mtp \
	-ngl 999 -fa on \
	-c 65536 --jinja --port 8083

7.8t/s

llama-server -hf unsloth/Qwen3.6-27B-MTP-GGUF:UD-Q6_K_XL \
  --spec-type draft-mtp --spec-draft-n-max 2 \
  -ngl 999 -fa on -c 65536 -np 1 --jinja --port 8083 \
  --alias qwen3.6-27b \
  --batch-size 2048 --ubatch-size 1024 \
  --cache-reuse 256 \
  --temp 0.6 --top-p 0.95 --top-k 20 --min-p 0.0 \
  --presence-penalty 0.0 --frequency-penalty 0.0 --repeat-penalty 1.0 \
  --chat-template-kwargs '{"preserve_thinking":true}'

4.5t/s

# https://unsloth.ai/docs/models/qwen3.6#mtp-qwen3.6-27b
llama-server -hf unsloth/Qwen3.6-27B-MTP-GGUF:UD-Q4_K_XL \
    --temp 0.6 \
    --top-p 0.95 \
    --top-k 20 \
    --min-p 0.00 \
    --spec-type draft-mtp --spec-draft-n-max 2 \
    --port 8083

36t/s (128k context)

llama-server -hf unsloth/Qwen3.6-35B-A3B-GGUF:UD-Q4_K_XL \
    --temp 0.6 \
    --top-p 0.95 \
    --top-k 20 \
    --min-p 0.00 \
    --port 8083

36t/s (128k context)

llama-server -hf unsloth/Qwen3.6-35B-A3B-GGUF:UD-Q4_K_M \
    --temp 0.6 \
    --top-p 0.95 \
    --top-k 20 \
    --min-p 0.00 \
    --port 8083

llama-server -hf unsloth/Qwen3.6-35B-A3B-NVFP4 \
  -ngl 99 \
  -fa on \
  -c 131072 \
  -ctk q5_0 \
  -ctv q4_0 \
  -b 2048 \
  -ub 1024 \
  --host 127.0.0.1 \
  --port 8080

I couldn't get the following models working:

llama-server -hf unsloth/Qwen3.6-35B-A3B-NVFP4 \
	-ngl 99 \
	-c 131072 \
	--port 8083

llama-server -hf unsloth/Qwen3.6-35B-A3B-NVFP4-Fast \
	-ngl 99 \
	-c 131072 \
	--port 8083

0.00.541.118 E get_hf_plan: no GGUF files found in repository unsloth/Qwen3.6-35B-A3B-NVFP4-Fast 0.00.541.123 I Available GGUF files: failed to download model from Hugging Face

---

LM Studio Comparison

For comparison I ran the raw qwen (not unsloth) model on LM Studio:

• qwen/Qwen3.6-35B-A3B-Q4_K_M.gguf (22GB)

• 256k context size

• 40.50 t/s

• 38 t/s second run

• Qwen3.6-27B-GGUF/Qwen3.6-27B-Q4_K_M.gguf

• 256k context size

• 10 t/s

tokens-persec_m1_max_64gb.jpg

I just ran the new Qwen 3.8 models from lmstudio against a prompt:

As a staff or principle software engineer with years of experience I need you to create an example list of 10 common bugs that show up in Go / Golang programs. Examples include concurrency bugs, nil-pointer exceptions, and array or slice allocations problems. Additional common bugs around mutexes and common stdlib usage should be included.

• https://huggingface.co/lmstudio-community/Qwen3.8-27B-GGUF (256k context window) = 8.91 t/sec, 10278 tokens
• https://huggingface.co/lmstudio-community/Qwen3.8-27B-MLX-4bit (256k context window) = 14.35 t/sec, 15478 tokens