← Back to context

Comment by raffraffraff

12 hours ago

M1 Max 64gb here, running Qwen 3.6 35b (unsloth 4bit) and it's plenty usable, and cheap used. Best bang for buck.

4 comments

raffraffraff

Reply
yieldcrv  9 hours ago

heyyy same, checkout kapitsa as the inference engine, gets 145+ tokens/sec

https://github.com/nace-ai/kapitsa

this machine just keeps getting better and better in the inference world. the memory bandwidth is too good, newer macbook pros aren't that much faster in memory bandwidth, many are slower. not enough to upgrade yet as they're mostly horizontal upgrades, the improvements in other areas are adding up like wifi bandwidth and protocol, NVMe speeds, webcam

if they do the M7 correctly: which is just everything their current models do plus way faster memory bandwidth and much greater memory possible 192GB+, then it will be a really great and timely upgrade

  • no-name-here  1 hour ago

    > improvements in other areas are adding up like … webcam

    Has the webcam changed or improved in the last couple years? It seems to still have the same webcam specs as the M4?

  • raffraffraff  4 hours ago

    Here are the numbers that got me buying a used M1 Max for LLM work.

        M1 Max 400GB/s
    M4 Pro 273GB/s
    M5 Pro 307GB/s

    Only Max are worthwhile for LLMs. An Ultra is just two Mac chips on a die, but I think they're just used in Mac Studio.

    I'll check it that inference engine! Thanks!

    Edit: it doesn't give numbers for unsloth/Qwen-3.6-35B-A3B 4bit, but I'm using mlx-serve right now and getting petty good numbers (and it took a while to tweak everything properly!)

    • yieldcrv  4 hours ago

      it has just a few models it can download, it's really qwen and apple silicon pilled

      it has unsloth on the list of 3.6 35b variants only on the 2-bit unsloth UD-IQ2_XXS GGUF variant

      I'd say let it down 4-bit Qwen 3.6 35b again and see what how it performs in both benchmarks and your preferred conversational and agentic harness

      it's Kflash verifier is competing against MTP and Dflash2, thats where the additional novelty is