I like the technique described here around distillation to recover from quantization, but I don't understand why we keep performing lossy compression on LLMs then using benchmarks that were nearly saturated before post-training to measure the effects.
You could erase the gains from literally half the compute going into some of these recent models and barely make a dent in MMLU-Pro and GPQA-D.
I'm still kind of surprised that people are targeting edge deployment of MoE models. By definition they optimize for computation cost at the expense of memory efficiency. We generally need the opposite on the edge.
I'm hoping to see more work in the other direction with cyclic/looped transformers and other memory dense approaches.
Hi Guanming/Bill. Would love to chat about what you're doing for actually running the models. I'm in a similar space, speeding up the `docker pull` component of inference deployment on edge devices (among other things!) If you're interested, shoot me an email at kyle@clipper.dev
Through my estimations, based on Bonsai's parameters/GB ratio, if one model were to have this ratio and Gemma4:12b's size, it would have the nice number of 54.125b parameters (that could run on 16GB of RAM). Is there any organization attempting something of this kind?
Have you run ablations on the actual effect/impact of on-policy distillation on contributing to the performance ? Just Curious ! As Unsloth based mixed quantisation methods on MoE models are widely used with great community rep.
For those too lazy to watch someone talk on video for ages to make a point:
The link is to a famous YouTuber called PewDiePie and he uses a local LLM to parse his email, to save time with that. They have an autoreply system and get notified about urgent matters.
Have you benchmarked against other 3-bit dynamic quants like Unsloth? I am sorry but this framing against a full precision, newer, smaller MoE just seems misleading. Also, Gemma-4-26B-A4B is not the SOTA for edge. Even at launch, that would be the 31B.
I can't find it. Can you state your performance versus comparable 3-bit quantization from Unsloth/Bartowski? Edit: I appreciate that you seem to have open-sourced the quantization pipeline. This is not to question your work, but to understand where the outputs stand relative to the SoTA for quantization.
I like the technique described here around distillation to recover from quantization, but I don't understand why we keep performing lossy compression on LLMs then using benchmarks that were nearly saturated before post-training to measure the effects.
You could erase the gains from literally half the compute going into some of these recent models and barely make a dent in MMLU-Pro and GPQA-D.
As an aside, General Instinct sounds like the name I'd give a megacorp in one of my cyberpunk ttrpg campaigns.
In cyberpunk ad world, after the campaign the whole town is littered with GI tracts.
I'm still kind of surprised that people are targeting edge deployment of MoE models. By definition they optimize for computation cost at the expense of memory efficiency. We generally need the opposite on the edge.
I'm hoping to see more work in the other direction with cyclic/looped transformers and other memory dense approaches.
[flagged]
Hi Guanming/Bill. Would love to chat about what you're doing for actually running the models. I'm in a similar space, speeding up the `docker pull` component of inference deployment on edge devices (among other things!) If you're interested, shoot me an email at kyle@clipper.dev
Sorry if this is somewhat off-topic:
Through my estimations, based on Bonsai's parameters/GB ratio, if one model were to have this ratio and Gemma4:12b's size, it would have the nice number of 54.125b parameters (that could run on 16GB of RAM). Is there any organization attempting something of this kind?
Yes Google. They just released their Gemma 4 12b quant.
[dead]
Have you run ablations on the actual effect/impact of on-policy distillation on contributing to the performance ? Just Curious ! As Unsloth based mixed quantisation methods on MoE models are widely used with great community rep.
You've likely heard about this - he'd probably like to talk to you and might potentially give you some good PR.
https://www.youtube.com/watch?v=rAzT5lcezPs&t=467s
For those too lazy to watch someone talk on video for ages to make a point:
The link is to a famous YouTuber called PewDiePie and he uses a local LLM to parse his email, to save time with that. They have an autoreply system and get notified about urgent matters.
Thanks for sharing! I'd love to chat with him. Would you be open to introducing us? :)
I assume PewDiePie runs something like DeepSeek 4 Flash on that rig.
Have you benchmarked against other 3-bit dynamic quants like Unsloth? I am sorry but this framing against a full precision, newer, smaller MoE just seems misleading. Also, Gemma-4-26B-A4B is not the SOTA for edge. Even at launch, that would be the 31B.
Yes I did, with other SOTA quant methods like HQQ, AWQ etc. You can find more info in our blog :) https://general-instinct.com/blog/frontier-moe-sub-4-bit
I can't find it. Can you state your performance versus comparable 3-bit quantization from Unsloth/Bartowski? Edit: I appreciate that you seem to have open-sourced the quantization pipeline. This is not to question your work, but to understand where the outputs stand relative to the SoTA for quantization.
How many watts? How does it effect power envelope?
[flagged]