Loading activity…
@primeintellect
Open Superintelligence Factory.
AI
Prime Intellect says revisions to AutomationBench’s tool descriptions, tasks and graders drastically changed model scores.
AI
PrimeIntellect says Prime Inference's 96.9% cache hit rate brings effective input costs to $0.024 per million tokens.
AI
PrimeIntellect also claims up to 34% faster hosted training with prime-rl in tests on two Qwen3 models.
XAI
Prime Intellect says its Qwen3-235B-A22B SFT ran 1.15× faster than torchtitan on the same nodes, with the same memory use.
AI
PrimeIntellect says IPO is tailored for stability in asynchronous reinforcement learning.
AI
Prime Intellect's event listing plans a research talk and discussion but marks the event full with a waitlist.
AI
A user says Moondream 3 is finally available there and links to a video.
AI
PrimeIntellect says Sword Health RL-post-trained GLM-5.3 into Dawn 1, claiming it is safer than its base.
AI
PrimeIntellect says it built 2,500 hard enterprise tasks with NVIDIA, modeled on AutomationBench.
Technology
NVIDIA says the model combines reasoning and tool use with text, image, document and video understanding.
YouTubeAI
Nvidia says the open model combines reasoning and tool use with text, image, document and video understanding.
YouTubeAI
Ai2 says it is co-hosting the free event with Hugging Face at The Midway in San Francisco.
Technology
A post links to OpenTogether and says its author is hearing there are “a lot of surprises.”
AI
Prime Intellect says a 4B TimeRLM trained on fresh AnomalyXL data recursively called sub-agents to locate anomalies.
AI
Prime Intellect says it is bringing Moondream’s open vision models and inference optimizations into its stack.
XAI
PrimeIntellect says the rewrite reaches usable input about 13 times faster and uses 83% less startup memory than before.
AI
PrimeIntellect says messages cost tokens and lose information, while agents must coordinate to avoid duplicating work.
AI
A post says its persistent memory swarm would preserve and build on work in a shared Git repository.
XAI
GoodfireAI says participants will build models using a 150-million-cell atlas, with top teams’ hypotheses tested in Prima Mente’s lab.
AI
Goodfire says external red-teaming from @farairesearch found its monitor cut universal jailbreak success to 0%.
X
