catalog: add q2_0 quant and a Q2_0 board for Qwen3.8-Flash-Next

[shacortes]

What

Adds a q2_0 entry to QUANTS and puts it on qwen3-8-flash-next's quant list.

{ id: 'q2_0', label: 'Q2_0', bits: 2, format: 'GGUF' },

Why

Qwen3.8-Flash-Next is 176.9B parameters (llama.cpp reports 176.94 B loading it; the HF safetensors metadata reports 180.0B), and the smallest full-weight GGUF published for it is 61.9 GiB. Its existing boards are fp8, q4_k_m and q8_0, which start at roughly 104 GiB — so no single consumer or workstation card has a board to post against for this model.

The catalog had no sub-4-bit GGUF quant at all, so this is the first one. Placed at the head of the GGUF group by bit width, ids keeping their underscores per frontend/src/catalog/README.md.

Checks

npm --prefix frontend run catalog:validate
  ✅ The catalog validates: 46 hardware, 32 models, 16 quants, 7 runtimes.

npm --prefix frontend run typecheck
  (clean)

catalog:seed writes SQL to stdout rather than a file, so there is nothing to commit alongside; regenerating the seed migration is a maintainer step per the catalog README.

Follow-up

I have a measured Q2_0 result for this model on an Intel Arc Pro B70 ready to submit once this is merged and deployed — 28.15 tok/s decode / 453.66 tok/s prefill on stock llama.cpp SYCL, with expert tensors partially offloaded to system RAM. Sending it as a separate PR, as results/README.md asks.

[vercel]

@shacortes is attempting to deploy a commit to the Community Labs Team on Vercel.

A member of the Team first needs to authorize it.

[github-actions]

No result files to import.

Site / sign up · Workflow details and retry

[shacortes]

I'm just thinking of making a PR that actually adds several quant options to all of them at this point, as doing PR per result isn't the best thing. The problem is that all models have more or less the same quantizations, so it'd just clutter it up a bit. Is it not possible to make the quants global?

[jackwsmth]

The problem is that all models have more or less the same quantizations, so it'd just clutter it up a bit. Is it not possible to make the quants global?

I'll take a look into this today. I understand what you mean.

[jackwsmth]

Okay I've taken a look into it further; I think for now the duplication isn't really an issue.

However, if it becomes increasingly annoying for contributors to submit results because they have add new quantizations to the global catalog, then yes I think it might be worth setting up some already created per-model lists.

As more people contribute, I think the global catalog will end up filling up with the quants anyway, and newer contributors won't have to keep adding to the global quants set.

That's my opinion on the matter at least, but open to any feedback.