I got early access to @typesafeai’s Jev—the new “System One” model that doesn’t generate text.
I tested the live API.
Headline: 50 semantic judgments in 226 ms.
One state, one request, all answers together.
The parallelism looks real. 🧵
After co-inventing ChatGPT, I kept asking myself: why have superhuman chat models not led to AGI?
I’ve spent the last 2 years in stealth building a new way to train models (RLCD), and a new type of frontier AI model that we are releasing today: Jev
• 20-200x faster
• 40-400x
OpenAI has unlimited compute. They hear a Millennium problem might be getting solved:
Option A: Do nothing
→ Problem takes N more months
→ Public reaction: “Ugh, imagine having unlimited resources and just sitting on them. Disgusting.”
...
current model + harness benchmark for agentic work:
"[Airport Code 1]-[Airport Code 2]
[Date 1]-[Date 2] ± 1 day
1 Person
Mandatory domestic layover if no direct flights.
[Budget]
Anything available right now?"
Shocked with my initial tests:
@Alibaba_Qwen is the provider for this model on @OpenRouter
Avg decode: 52 tk/s
Right now, you can run a model better than:
>Opus 4.7(max)
>Sol 5.6(med)
>Opus 5(low)
>GPT T.5(xhigh)
fully offline ✅
as fast as the official API ✅
with near lossless intelligence performance
@Alibaba_Qwen 3.8 Flash Next on a single DGX Spark @ 50 tok/s C1, 256k ctx, 2,000 tok/s on cold prefill running @vllm_project
Agentic loop produced observed decode rates of 74–81 tok/s
Top-1 88.6%
Mean KLD 0.040
BF16 KV Cache
NVFP4 from @NVIDIAAIgithub.com/gitcommit90/qw…
