The newest community items on cmp.wiki, from November 10 to 13, 2026. The week followed Kate's GA100VFI write-up of November 10, which reports 80 GiB on nine of ten 8 GB cards after a hardware procedure with permanent fuse changes; the wiki keeps that result on Unlock internals. Each item gives its wiki status: Reported means one person's result, and Contested means members disagree.
Prices
- 8 GB card prices jumped (Reported): members paid about 2,800 USD on eBay on November 11, 2026, eBay asks of 5,000 and 7,000 USD were seen on November 12, and the cheapest 8 GB listing on eBay was 4,400 USD on November 13, per @froggao. Several orders were cancelled and relisted higher. Members disagree whether the 10 GB card now costs more or less than the 8 GB card. The Price log gained 28 rows; the week has its own page, Price history, November 2026.
Unlocks and hardware
- 96 GB on the 8 GB card (Contested): Kate says the sixth HBM stack on the cards she tried is a dummy die, while @cute_sc2 reports 2 of his 21 cards with every memory partition healthy (November 12 and 13, 2026). See Rumours and claims to verify.
- NVLink (Reported): @zorg33 reports chained NVLink working on 10 GB cards, with no bandwidth numbers yet (November 10, 2026), and members are working on a copy of the A100 bridge, held up by its edge connector. See NVLink.
- cmpunlocker: pull requests #67 and #68 were closed without merging on November 10, 2026 UTC. Four new pull requests from alexxy (November 13) fix the 64 GB BAR1 on every card of a multi-card host behind a PCIe switch, build 615.71.09 on Linux 7.3 and add a Gentoo mode; tested by their author on two 8 GB cards. See Resizable BAR and BAR1 64 GB.
Recipes and engines
- GLM 5.3 Flash on two stock 8 GB cards (Reported): @alexbrandes runs a 2.6 bpw EXL3 quant in vLLM over x4 links without P2P at a little over 70 tok/s decode and 2,500 tok/s prefill (November 11, 2026). See the new recipe page.
- Project Maya v1.1.58 to v1.1.61 (November 10 to 12) add a 170 GB quant of the full GLM-5.3 and bring DeepSeek decode on one NVIDIA card back to speed; not measured on CMP cards yet. See Engines.