So let me describe the workflow and perhaps I can tell you about how I go about it and what settings I use and what I like about Qwen3.5 and what I like about Gemma4.

Models:

• Qwen3.5-35B-A3B unsloth Q8_K_XL variant
• Gemma4-26B-A4B-IT unsloth Q4_K_XL variant
• One might ask "why different variants"...don't ask this. It sorta comes down to speed, but it also sorta comes down to "this is what I downloaded first and it worked so well that I stuck with it"...

Qwen3.5 Launch settings:

/opt/homebrew/bin/llama-server \
  --model ~/models/Qwen3.5-35B-A3B-UD-Q8_K_XL.gguf \
  --mmproj ~/models/Qwen3.5-35B-A3B-mmproj-F32.gguf \
  --ctx-size 131072 \
  --n-gpu-layers 999 \
  --threads 16 \
  --parallel 1 \
  --batch-size 1024 \
  --ubatch-size 1024 \
  --cache-type-k bf16 \
  --cache-type-v bf16 \
  --flash-attn on \
  --jinja \
  --temp 0.7 \
  --top-k 20 \
  --top-p 0.95 \
  --metrics \
  --mlock \
  --host 0.0.0.0 \
  --port 8092

Gemma4 Launch Settings:

/opt/homebrew/bin/llama-server \
  --model ~/models/gemma-4-26B-A4B-it-UD-Q5_K_XL.gguf \
  --mmproj ~/models/mmproj-gemma-4-26B-A4B-it-F32.gguf \
  --port 8081 \
  --ctx-size 262144 \
  --n-gpu-layers 999 \
  --threads 24 \
  --threads-batch 24 \
  --flash-attn on \
  --cache-type-k bf16 \
  --cache-type-v bf16 \
  --temperature 1.0 \
  --top-p 0.95 \
  --top-k 40 \
  --min-p 0.01 \
  --mlock

• One might also note "Hey, those aren't apples to apples settings!" and you'd be right...different models get different settings. This is where I put on my deal with it sunglasses. This is what I've found works best so far for me/us.

So the workflow is to paste in some heinously terrible handwritten document that looks similar to something like https://mtv-drupal-assets.s3.amazonaws.com/files/resources/chastellux-letter-vine-12.jpg but usually far worse quality / ink spills / smudges / faded writing / etc (note: that one is a modestly recognizable one from George Washington's Mt. Vernon collection and has nothing to do with me personally).

I use this prompt typically as my basic test prompt:

• Can you tell me what this image is and if it contains text, can you please take some time and try to transcribe it? Please add (?) where you are unsure/uncertain about a piece of text. Don't get stuck though if it's indecipherable - just mark it as (?) and move on.

Workflow (once the model is up and warm and all that):

• Paste in picture of terrible ~~cursed~~ cursive handwriting that someone committed to paper 100-300+ years ago.
• Paste in the above prompt.
• See what magic (or not) comes out.

Here's what I like about Qwen3.5-35B-A3B:

• It's a thinker. It will analyze every single line, character, ponder, go back and look again if it finds the same word or similar words but realizes it made a different transcription previously and then it will compare the two, decide if they're the same or not and then after some indeterminate amount of time (usually like 60-100 seconds), it will pop out an answer. I've watched it reason in real time and on my hardware (M3 Ultra Mac Studio 512GB w/ 80 GPU cores), it gets like 65tok/s and can output thousands of tokens in the thinking blocks before it pops out a final answer. That answer is usually pretty good (usually like 2-5 words incorrect on a 100 word document). However...it almost never uses the (?) uncertain tags. It's a confident sometimes incorrect model that kind of just ignores that part of the prompt.

Here's what I like about Gemma4-26B-A4B-IT:

• It's fast and thinks substantially less, but comes up with the same answer as Opus like 95%+ of the time. It still reasons quite a bit - but way less than Qwen3.5. Qwen will spend 70 seconds pondering over the letter T and Gemma won't. Gemma is happy to mark it with a (?) and let a human figure out whether that's a T or an I. This is good behavior in my opinion. It also will spit out an answer in 20-30 seconds instead of 70-100 seconds...and it's usually just as good or better than Qwen's answer. Gemma does not ignore the prompt and does the same quality (+/- a small percentage) as Qwen in far fewer tokens and far less time. 25 seconds vs 80 seconds is kind of a big deal for me especially if the output is about the same quality.