Select Your AI: Cloud vs My Mac, Graded Blind

Pick your fighter. I put six AI setups on a select screen, gave five of them the same job, and had a blind judge score the results. Two are Claude, in the cloud. Three are Qwen builds running on my own Mac. The sixth is a 233 GB DeepSeek, the final boss, on a Mac with 128 GB of memory.

The job

On November 4, 2026 the five graded fighters got one task written to my Shorts channel's rules: write a ~120-word Short script from a GitHub Changelog post (November 2, 2026). The rules: a hook in the first 1 to 2 seconds, only facts from the source, no emojis, no income claims and a fixed closing line.

The judge was Codex (OpenAI). It saw five scripts labelled A to E with the names hidden, and scored each one out of 50 on hook, accuracy, clarity, rules and whether it would ship it.

The roster

Fighter	Where	Blind score (/50)	Speed on the script job
Claude Sonnet 5.5	Cloud	45	13.9-18.5 s per script (3 runs)
Claude Opus 5.5	Cloud	44	16 s per script (1 run)
Qwen3.8-27B GGUF Q4_K_M (16.5 GB)	My Mac	38	33 tok/s
Qwen3.8-27B MLX 4-bit (16.1 GB)	My Mac	35	30 tok/s
Qwen3.8-27B MLX 8-bit (29.5 GB)	My Mac	27	17 tok/s
DeepSeek V4.1 Flash IQ1_S (233 GB)	My Mac	Not graded	0.42 tok/s, one-sentence test, busy Mac

How to read it:

• The Mac is an M5 Max with 128 GB. The Qwen builds ran in LM Studio. DeepSeek ran in llama.cpp after LM Studio refused to load it.
• Qwen speeds are quiet-Mac medians of 3 runs on the script job, thinking off. The Claude times are wall time per script, including the CLI start and the network, so there's no tokens-a-second number for them.
• Each model wrote one script. The order within the two Claudes and within the three Qwen builds is noise, and Sonnet's top score is one point ahead of Opus.
• The MLX 4-bit build has a special move: with thinking on, it hit the 8k-token cap and gave no answer.
• Chip power from the macOS sensors (CPU+GPU+ANE, a rough average on a busy Mac): about 39 W for the GGUF build and 32 W for MLX 4-bit.

The final boss

DeepSeek V4.1 Flash in its IQ1_S build is a 233 GB file. My Mac has 128 GB of memory. It couldn't run the script job under the test's guards, so it isn't in the blind test. On a one-sentence test on a busy Mac it wrote 23 tokens at 0.42 tokens a second, with the chip at about 49 W. The full run is in Episode 07 on the channel.

My pick

For writing scripts on a Mac, I picked the Qwen GGUF build: the fastest of my three blind-test builds and the top local score, 38 (one script each). On my own weights (70% blind score, 30% runs on my Mac) it came out at 83%, ahead of MLX 4-bit at 79% and MLX 8-bit at 68%. Those weights are my choice, so treat the percentages as my opinion.

The model runs on my Mac, and LM Studio's Free plan runs local models at $0. Both Claudes still scored higher: 45 and 44 against 38.

Try it yourself

1. Install LM Studio from lmstudio.ai. The Free plan runs local models on your machine, with llama.cpp and MLX under the hood.
2. Give the same real task to each model you're thinking about, local and cloud.
3. Hide the names and have someone else score the answers: a colleague or a different AI.
4. Score quality, speed, cost and privacy, then pick with your own weights.

Here's how you get paid

My opinion, no income guaranteed: run the same select screen for a client. Blind-test a few AI models on their own work, score them on quality, speed, cost and privacy, then set up the one they pick, local or cloud.

• Test. Their real task, a few models, names hidden.
• Score. The same four stats for every model.
• Set up. Install and configure their pick.