vLLM hits 954K tokens/s per GPU on AgentX
Under DeepSeek’s off-peak pricing, the open-source serving engine cleared $38.96 per GPU-hour and 954K tokens/s of combined throughput on real-world agentic workloads.
Beyond matching DeepInfer’s headline numbers, vLLM put itself through the SemiAnalysis AgentX suite of real-world agentic workloads. At DeepSeek’s off-peak pricing, the engine delivered $38.96 in revenue per GPU-hour, with 954K tokens/s per GPU of combined cached, input and output throughput at 196 tokens/s of generation. The result turns inference economics from a paper benchmark into a live revenue figure — and sharpens the argument that serving infrastructure, not just model quality, is where the margin now lives.
SGLang fully matches the DeepInfer benchmark
DeepInfer set the bar; SGLang pushed it higher. The project now matches DeepInfer on both prefill and decode, reaching up to 1.04× prefill throughput on 4× GB300 and 4× B300, and up to a 7.4% decode gain on 64× GB300.
“Managing how much you guide the AI is the #0 skill I look for.”
Guillermo Rauch, Vercel CEO
Kling 4.0 API goes live on Atlas Cloud
Atlas Cloud now serves the full Kling V4 API from Kuaishou, bringing 4K HDR video, 21:9 ultra-wide frames, multi-keyframe control and multilingual dialogue into production. Omni Reference accepts up to 15 references per call.
vLLM matches DeepInfer on DeepSeek-V4.1-Flash
vLLM matched DeepSeek’s own inference engine across decode configurations with batch sizes from 8 to 256, running DEP32 against DeepInfer’s EP64DP64. DeepInfer was originally built on vLLM and now integrates DeepSeek’s open-source kernels.
Claude Code author Thariq confirms Anthropic exit
Thariq, the author of Claude Code, has officially announced his departure from Anthropic — a closely watched personnel signal in the agent-tooling space.
Jensen Huang pitches a European AI economy at GTC Berlin
At GTC Europe, Huang argued that Europe’s generations of science, engineering and manufacturing expertise can now power AI of its own — and urged a shared push to put AI to work.
Seedance 2.5 renders native 4K multi-scene transitions
Higgsfield used a Blender blockout as reference and Seedance 2.5 produced continuous transitions across a dozen scenes — from a medieval village to a neon city in the rain — at native 4K.
OpenAI and Azure scale to 1.2B users
OpenAI Devs and Azure Cosmos DB engineers say they jointly support 1.2 billion users across roughly 40 regions with single-digit-millisecond latency.
Computer launches open-source skill /impeccable
The open-source skill, built by Paul Bakaus, generates interfaces that don’t read as AI slop. Each user gets 5,000 credits to try it.
Enact raises $30M pre-seed for dexterous robots
Enact emerges from stealth to master dexterity through force and touch, keeping vision and language as core inputs.
Flow-map LMs can act as speculative-decode drafters
Flow-map language models capture token dependencies in one to two passes, making them natural drafters for speculative decoding.
Dropping local correlations sharpens image representations
A new method abstracts away local detail to focus on weaker long-range correlations, yielding better representations and 128/256 ImageNet generations.
Take “out there” AI-security warnings seriously
Ethan Mollick argues the labs’ claims about AI transforming cybersecurity deserve more weight than they have received.
Anthropic accused of removing Claude’s reverse-engineering
Circulating reports claim Anthropic quietly shut down Claude’s REA capability, reigniting the local-models debate.
OpenAI hardens its research training environment
OpenAI says it has substantially hardened the security of its research training environment over recent months and removed related risks.
Researcher recalls an LLM training-poisoning ordeal
Shoufa Chen recounts months of investigation, confrontation and doubt in 2024 as a direct victim of a training-poisoning incident.
“The Moore’s law of generative AI is consistent: quality, resolution and control double about every year.”
Cristóbal Valenzuela
Block releases Dessa full-duplex speech model
A training method driven by simple human curiosity yields a strikingly realistic speech model — and a template for more novel approaches ahead.
Kling 4 nails a four-character face swap on the first try
A near-perfect first-attempt four-character video swap, and it runs faster and cheaper than Seedance 2.5.
Gemini 4 typesets complex chemical structures in LaTeX
A nanoputian macromolecule rendered in ChemFig, complete with mass and molecular weight, hints at scientific document tooling.
Seedance 2.5 shows what native 4K skin detail looks like
Higgsfield’s side-by-side argues that generating in 4K directly — rather than upscaling afterward — preserves the fine texture of skin, a subtle but telling quality signal for video models.
CyberLink integrates Kling 4.0 into its imaging suite
PowerDirector, PhotoDirector and MyEdit all pick up Kling 4.0.
How JEPA picks its world-model features
A thread asks whether JEPA keeps hard-to-predict features like a traffic light.
Real-world usability beats ±2 points on AA
At the top end, some models “hit” and some “miss” beyond the benchmark.
Teacher forcing turns time iteration into depth iteration
François Fleuret explains why Transformers keep GPUs happy.
A B300 that could fund the whole lab
Graham Neubig jokes about serving DeepSeek-V4.1-Flash tokens to fund CMU research.
Don’t ignore single-agent systems
A shared blog argues single agents beat multi-agent setups on key axes.
Grok Bot keeps sensitive info in a vault
Musk says Grok Bot will store sensitive information in a dedicated vault.
Kimi K3.1 Pro questioned over review-boosting
The model is strong on its own, critics say, but trying to game reviews dragged the launch through the mud.
Kimi K3.1 Pro scores non-trivial on Mazebench
The first open-weight model to post a meaningful Mazebench result.
DeepSeek V4.1 Flash finds 54 vulns for $7.53
Seven more findings than the comparison model, at roughly 16.6× lower cost.
Claude Code author Thariq confirms Anthropic exit
op7418 says Claude Code author Thariq has officially announced his departure from Anthropic, seen as a significant personnel signal.
Jensen Huang discusses European AI economy at GTC Berlin
At GTC Europe, Jensen Huang called for building a European AI economy and putting AI into actual production.
Seedance 2.5 native 4K takes on complex transitions
Higgsfield used a Blender layout as reference to show Seedance 2.5 generating multi-scene continuous transitions at 4K.
China requires institutions to draft AI unemployment plans
A tweet says China has required government agencies to prepare plans for large-scale AI unemployment and called on the US to follow suit.
Block releases Dessa full-duplex speech model
Linus says Block's Dessa speech model is realistic and based on a new training method driven by human curiosity, potentially spawning more new tools.
Kling 4 four-character video face swap near perfect on first try
venturetwins says Kling 4's first attempt at four-character video replacement was nearly perfect, and faster and cheaper than Seedance 2.5.
Generative AI quality roughly doubles every year
c_valenzuelab argues that generative AI quality, resolution and controllability roughly double each year.
Vercel CEO: managing AI guidance level is core skill
rauchg argues both over-guidance and under-guidance cause problems, and managing this 'superpower' is the top skill.
Gemini 4 generates complex chemical LaTeX structures
fofrAI shows Gemini 4 using ChemFig to typeset nanokid polymer structures and provides quality and molecular weight.