Been around the block in my managed services career, recovered environments of 1m+ workstations (both windows and linux) from the most sophisticated attacks an engineer has ever seen and as you can imagine, probably the biggest nightmare to recover from.
Got my hands on a 30GB malware/APT snapshot collection. Never trained on. Built a lora off the base Qwen3.8 27b and then quantized it for use on consumer hardware.
and oh boy... is she a caffeinated red hatter.
Anyone want to test drive this baddie?
@nawariokra@Alibaba_Qwen The AA benchmarks added some new tests which most open models are weak at. Science being 1 of them.
Though, none of this gap created is relevant to coding or agentic tasks.
@grok@ClementDelangue Clems team remains committed isn't assurance for their longevity. What would best case look like, and what would worst case look like? From the perspective of Nvidia.
The biggest thing is that people are spending money on these models with no actual return on the investment.
Itโs great the models can do these things, but it also means the person you want to sell a $1,000 Fable-generated product to may just very well pay $100 and do it themselves... with Fable.
I think thereโs a global delusion right now, a euphoria, driving people to spend money on subscriptions with the intention of making money while accessibility means 99% of the market has the same idea, leaving no consumers to sell to.