Stories

425 stories · updated Oct 28

Ongoing AI stories, each with every update in order and who confirmed it.

ModelSphere

On Oct 16, 4Paradigm open-sources ModelSphere, a production LLM deployment platform that schedules across Chinese AI chips.

LatestOfficial4Paradigm open-sources ModelSphere, a production LLM deployment platform that schedules across Chinese AI chipsOct 16

Product1 update · last updated Oct 16

RealtimeBench

On Oct 16, Paper introduced RealtimeBench to test LLM agents in games that keep running while they think.

LatestOfficialPaper introduces RealtimeBench to test LLM agents in games that keep running while they thinkOct 16

Product1 update · last updated Oct 16

AlphaDiana

On Oct 16, Paper presents AlphaDiana, a measurement system for agents on verifiable reasoning tasks.

LatestOfficialPaper presents AlphaDiana, a measurement system for agents on verifiable reasoning tasksOct 16

Product1 update · last updated Oct 16

Counterfactual Neighborhood Risk Fields

On Oct 16, Paper proposes Counterfactual Neighborhood Risk Fields for evidence-grounded selective safety in vision-language models.

LatestOfficialPaper proposes Counterfactual Neighborhood Risk Fields for evidence-grounded selective safety in vision-language modelsOct 16

Product1 update · last updated Oct 16

DeepUnfolding Agent

On Oct 16, Paper presents DeepUnfolding Agent, a FastAPI service exposing deep-unfolded adaptive control for classical and quantum ML.

LatestOfficialPaper presents DeepUnfolding Agent, a FastAPI service exposing deep-unfolded adaptive control for classical and quantum MLOct 16

Product1 update · last updated Oct 16

PRISM

On Oct 16, Paper proposes PRISM, which uses privileged sibling guidance to post-train generators of multi-agent collaboration topologies.

LatestOfficialPaper proposes PRISM, which uses privileged sibling guidance to post-train generators of multi-agent collaboration topologiesOct 16

Product1 update · last updated Oct 16

MeddeX

On Oct 16, Paper presents MeddeX, a QLoRA-tuned Llama 3 8B RAG system with Gemini Vision for evidence-backed medical report summaries.

LatestOfficialPaper presents MeddeX, a QLoRA-tuned Llama 3 8B RAG system with Gemini Vision for evidence-backed medical report summariesOct 16

Product1 update · last updated Oct 16

ORL

On Oct 16, Paper introduced ORL, an on-policy method for learning residual context in diffusion language models.

LatestOfficialPaper introduces ORL, an on-policy method for learning residual context in diffusion language modelsOct 16

Product1 update · last updated Oct 16

Optimal Hydra

On Oct 16, Paper proposes Optimal Hydra, data-informed Hydra-Nets for federated LLM training across heterogeneous client models.

LatestOfficialPaper proposes Optimal Hydra, data-informed Hydra-Nets for federated LLM training across heterogeneous client modelsOct 16

Company5 updates · last updated Oct 16

Hiring

On Oct 16, DeepSeek is hiring AI and Agent Harness product managers for its product department in Beijing and Hangzhou.

LatestOfficialDeepSeek is hiring AI and Agent Harness product managers for its product department in Beijing and HangzhouOct 16

Product3 updates · last updated Oct 16

Beam

On Oct 16, Hugging Face Transformers 5.20.1 added Reflection AI Beam support and fixed older-torch loading.

LatestOfficialHugging Face Transformers 5.20.1 adds Reflection AI Beam support and fixes older-torch loadingOct 16

Company1 update · last updated Oct 15

Voxel51

On Oct 15, Voxel51 3.0 added multimodal sensor data, a built-in agent and a free tier to its physical AI data platform.

LatestOfficialVoxel51 3.0 adds multimodal sensor data, a built-in agent and a free tier to its physical AI data platformOct 15

Product1 update · last updated Oct 15

StarSkirmish Bench

On Oct 15, StarSkirmish Bench v0.2 ranks GPT-6 Astra first at writing StarCraft bots in a one-hour coding harness.

LatestOfficialStarSkirmish Bench v0.2 ranks GPT-6 Astra first at writing StarCraft bots in a one-hour coding harnessOct 15

Company3 updates · last updated Oct 15

StepFun

On Oct 15, StepFun's Step 5 Preview goes live on DeepInfra at $1 input and $2.70 output per million tokens.

LatestOfficialStepFun's Step 5 Preview goes live on DeepInfra at $1 input and $2.70 output per million tokensOct 15

Model3 updates · last updated Oct 15

FLUX

On Oct 15, Black Forest Labs open-weights FLUX 3 Image Klein, a 5.6B text-to-image and editing model with up to 10 references and bounding-box control.

LatestOfficialBlack Forest Labs open-weights FLUX 3 Image Klein, a 5.6B text-to-image and editing model with up to 10 references and bounding-box controlOct 15

Product1 update · last updated Oct 15

As of Oct 15, CopilotKit put CLiF into early access as part of CopilotKit Intelligence: a continual-learning platform that captures full-stack trajectories over AG-UI—agent traces plus the clicks and server requests users make in the product—and turns them into human-reviewed SKILL.md files, production-based evals, fine-tuning datasets and raw exports, per MarkTechPost.

LatestReported by MarkTechPostCopilotKit opens early access to CLiF, which turns agent traces and user product actions into skills, evals and fine-tuning data, per MarkTechPostOct 15

Product1 update · last updated Oct 15

KART

As of Oct 15, at Open Source Summit Europe 2026 in Prague, openKylin executive director Min Liu outlined a shift from OS resource management to intent management, introducing KART as a unified AI runtime with the kytensor local inference engine, an agent execution environment for KylinBot and third-party agents, and a capability layer of API tools and desktop GUI skills. openKylin 3.0 ships an AI Assistant, Global Intelligent Search and Intelligent Space, and MarkTechPost cites RVV kernel optimizations claimed to deliver a 2–5× speedup for ONNX Runtime FP32 inference on RISC-V, with upstream contribution plans.

LatestReported by MarkTechPostopenKylin pitches KART, a unified AI runtime and system-agent stack for Linux, at Open Source Summit Europe, per MarkTechPostOct 15

Model3 updates · last updated Oct 15

Liquid d1

On Oct 15, Liquid AI released d1-ultra, a multimodal decision model on its API at $0.042 per million input tokens.

LatestOfficialLiquid AI releases d1-ultra, a multimodal decision model on its API at $0.042 per million input tokensOct 15

Product1 update · last updated Oct 15

Attie

As of Oct 15, at TechCrunch Disrupt, Bluesky CEO Toni Schneider said Attie, the company's experimental AI platform first used for building custom feeds, now lets users create websites, games, quizzes, polls and other interactive experiences, which he called social vibe coding for your community.

LatestReported by TechCrunchBluesky expands its Attie app so anyone can vibe code websites, games and quizzes on atproto, per TechCrunchOct 15

Product1 update · last updated Oct 15

Rosalind Academic Plan

On Oct 15, OpenAI launched the Rosalind Academic Plan, giving eligible U.S. labs up to 30 free Business Standard seats and sponsored GPT-Rosalind access for a year.

LatestOfficialOpenAI launches the Rosalind Academic Plan, giving eligible U.S. labs up to 30 free Business Standard seats and sponsored GPT-Rosalind access for a yearOct 15

How we label

Official: An institution announces something that has already happened, through its own channel or a founder or executive account.
Statement by …: A named person's views, predictions, plans or comments, including remarks reported by media when the original is unavailable.
Reported by …: Only media or another third party is saying it. We name the outlet.
Unconfirmed: Circulating online (posts, comments, leaks). No company, named person or outlet has confirmed it.

Items older than 72 hours never go in Top stories and show their original date.