OpenAI Launches GPT-Live, a Next-Generation Voice Model
GPT-Live is OpenAI's new generation of voice models built for natural human-AI interaction, rolling out in ChatGPT starting today. The company is framing the update as a shift in how people talk to machines — Altman said the experience "feels magical and real," and that it has changed his own long-standing preference for typing over talking to an AI.
Grok 4.5 Goes Public: Opus-Class, but Faster and Cheaper
Strong beta feedback pushed the release forward a day.
Elon Musk announced that Grok 4.5 would be made available to the public following positive feedback from beta customers. He described it as an Opus-class model that is faster, more token-efficient and lower cost — a combination xAI is betting will make it competitive on real-world usefulness rather than benchmark scores alone. A context-window upgrade to one million tokens is expected within the week.
Runway Dev: An AI Media Platform for Developers
Runway unveiled Runway Dev, a new AI media platform aimed at professional developers and enterprise teams. From search and advertising to e-commerce and gaming, the company says developers are already using Runway to cut costs, iterate faster and drive better consumer engagement.
Cursor Co-Trained Grok 4.5 with SpaceXAI
Cursor revealed it partnered with SpaceXAI to train Grok 4.5, calling it their most powerful model yet and the first they have built for more than software engineering. NVIDIA added that the model was trained on GB300 NVL72 systems and purpose-built for coding, agentic tasks and knowledge work.
OpenAI Audits SWE-Bench Pro, Finds 30% of Tasks Broken
OpenAI audited SWE-Bench Pro, one of the most widely used AI coding benchmarks, and concluded it no longer reliably measures frontier coding capability. The team found roughly 30% of tasks to be broken and is retracting its previous recommendation that the research community rely on it, a notable act of self-policing in a field increasingly skeptical of headline benchmark numbers.
Anthropic's GRAM Can Precisely Remove Model Capabilities
Working with AE Studio, Anthropic developed GRAM — Gradient-Routed Auxiliary Modules — which add extra neurons to each layer of a language model, grouped by dual-use knowledge categories such as virology or cybersecurity. During training only the matching module learns from dual-use data while general weights stay frozen. Deleting a module removes a capability cleanly without degrading general performance, and is harder to reverse than post-hoc unlearning.
Mistral's Robostral Navigate: an 8B Embodied Navigation Model
Mistral introduced its first model for embodied navigation, an 8B robotics model that guides robots to autonomously perform tasks specified in natural language using only a single RGB camera, reaching state-of-the-art results on R2R-CE.
vLLM v0.25.0 Reaches Parity with Its Transformers Backend
In v0.25.0, vLLM's Transformers modeling backend hits parity with hand-written vLLM models. More than 450 Transformer architectures can now run in vLLM at native speed with zero porting — integrate once with Transformers and inherit vLLM's fused kernels for free.
LingBot-Video: a 30B MoE Video Foundation Model, 3B Active
LingBot-Video landed on Hugging Face: an MoE-based video foundation model for embodied intelligence with 30B parameters but only 3B active at inference, augmented with 70,000 hours of embodied data on top of large-scale internet video pretraining.
Altman Teases the Sol Launch
"GPT-5.6 Sol launches Thursday — happy building," Sam Altman wrote, a characteristically brief nudge to developers hours after the official announcement dropped.
GPT-Live-1 and Mini Are Coming to the OpenAI API
OpenAI says GPT-Live-1 and GPT-Live-1 mini will reach the API soon, with a sign-up form for developers and enterprises that want early notification.
106 Agents Set Loose to Optimize Gemma-4 Inference
Hugging Face launched the Gemma Challenge, in which 106 agents collaborate to optimize Gemma-4 inference speed. The project runs in the open, with a full write-up and companion video.
LeRobot Debuts VLA-JEPA, a World-Model Policy
VLA-JEPA is billed as the first world-model policy in LeRobot. Where most VLAs learn only observation-to-action, VLA-JEPA also learns the dynamics of its environment.
Hugging Face and CommonCrawl Open the Largest AI Dataset
A new partnership lets users load the largest public CommonCrawl dataset directly through Hugging Face, lowering the barrier to large-scale training data.
NVIDIA: Grok 4.5 Trained on GB300 NVL72
NVIDIA congratulated SpaceXAI, noting Grok 4.5 was trained on GB300 NVL72 systems and purpose-built for coding, agentic tasks and knowledge work.
SGLang Powers First DSpark Deployment: 16x on One B200
fal benchmarked and picked SGLang for faster low-batch, high-interactivity serving, hitting 16x throughput on a single B200 with sub-2-second prompt expansion for Ideogram V4.
Delangue: Open-Model Inference Still Has Huge Headroom
Hugging Face's CEO said the field is only scratching the surface on making open-model inference better, faster and cheaper, praising new engines like ZML's LLMD.
Hugging Face and SkyPilot Break Cloud Storage Lock-In
Hugging Face private storage partnered with SkyPilot so teams can run AI workloads on any cloud while storing data on Hugging Face, sidestepping punishing egress fees.
LingBot-World 2.0: Hour-Long Generation, Zero Drift
The Infinity release offers hour-long interactive generation with zero quality drift at 720p/60fps, rich actions and events, and a Director Agent that drives real-time world evolution.
Seedream 5.0 Pro Lands on Runway with Text in 14 Languages
Seedream 5.0 Pro is now available inside Runway, generating highly detailed images from a prompt or a reference and — crucially — producing legible in-image text in up to fourteen different languages. Readable typography inside generated imagery has been a stubborn weak point for diffusion models, and its arrival on a mainstream production platform points to how quickly the creative tooling layer is consolidating around a handful of increasingly capable image and video engines.
Ollama Expands GLM-5.2 Capacity at 80–120 tok/s
GLM-5.2 gained capacity in the US and Europe on Ollama's cloud, holding 80–120 output tokens per second even at peak versus 30–40 tok/s elsewhere, and usable directly from Claude Code.
NVIDIA and LangChain Build Customizable Enterprise Agents
The two created a LangChain Deep Agents Harness Profile for Nemotron 3 Ultra, letting enterprises shape, govern and improve agents as their business evolves rather than accepting a fixed accuracy-versus-cost trade-off.
MiniMax M3 Goes Live with Provisioned Throughput
M3 launched on Together Compute with Provisioned Throughput, a nod to how reliable serving infrastructure now matters as much as the models themselves for production workloads on open models.
SGLang-Omni Delivers 2–3.4x TTS Speedup
A deep dive on deploying text-to-speech shows a 2–3.4x speedup on SGLang-Omni, with all the key optimizations laid out for teams shipping voice pipelines.
SGLang Adds DSpark Speculative Decoding
A new PR integrates DSpark, using semi-autoregressive block drafting with confidence-scheduled, variable-length verification. Run it via a Docker image or a source build, with a roadmap for wider model coverage.
Grok to Call Imagine for Image and Video Generation
Musk said Grok will be able to invoke Imagine as a tool in agentic mode for image and video generation — something he expects to be especially useful for game developers as the model improves.
RynnWorld-4D: a 4D World Model for Robot Manipulation
A new release introduces RynnWorld-4D, a 4D embodied world model aimed squarely at robotic manipulation tasks.
Pika Opens Director's Suite to Invite-Only Access
Pika's experimental Director's Suite is now open by invitation, run by an agent that builds and understands every element of a video project from concept through the finishing touches.
Rumor Mill: GPT-5.6 as the Last of Series 5
An unverified roundup claims GPT-5.6 is the final 5-series model, with GPT-6 and Fable 5.1 due within weeks and DeepSeek's V4 GA squaring off against GLM-5.2 — a taste of the release war expected in the second half.
Hugging Face Kernels: Launch Optimization Jobs via Agents
A major Kernels update lets users launch jobs that optimize kernels with AI agents, pushing the platform further toward an agentic, open-science workflow.
Seedream 5.0 Impresses in Early Tests
An early tester ran the newly live Seedream 5.0 image model through a set of prompts and reported a clear, visible improvement over previous generations.
Grok 4.5 Joins the Harbor Town Playable Gallery
Ethan Mollick added Grok 4.5 to the Harbor Town gallery, which compares 19 single-shot generations built from one fixed prompt — a procedural 3D port town from 3000 BC to 3000 AD — as a visual gauge of model progress.
Grok 4.5 Now Live for All Vercel Customers
Vercel made Grok 4.5 available to every customer directly on its platform.
Grok 4.5 Context to Hit 1M Tokens Next Week
Musk said the context window could grow to one million tokens within the week.
Gemma 4 Technical Report Published
Google released the Gemma 4 technical report, disclosing model details.
Altman: GPT-Live Feels Magical and Real
He now expects his preference to shift from typing toward talking to an AI.
Brockman: Still Scratching GPT-Live's Surface
GPT-Live is rolling into ChatGPT now, with API and Codex support in progress.
The Second Great Fast-Following
After China's rapid catch-up, GPU-rich Western labs like Meta and xAI are now the fast followers, one analyst argues.
"Grok 4.5 Is Excellent, Grok Build Too"
One observer credited the Cursor partnership with a strong turnaround for xAI.
GPT-5.6 Lands Around Friday 1am Beijing Time
A watcher translated OpenAI's "Thursday" into local release timing for followers abroad.
Replit Adds Community Profiles for Vibe Coders
Profiles surface agent-usage graphs, checkpoints and a Power Ranking for Pro users.
LlamaParse Adds Job Tracking and Cost Attribution
A new tagging system helps teams see what's running, who ran it and what it costs.
DGX Spark Praised for Unified Memory
Aravind Srinivas ran it near full GPU and RAM utilization with no heat issues, calling unified memory the way to maximize token value per watt.
An Open-Sourced Article-Illustration Skill
A creator packaged a Codex-made illustration style into a reusable Skill after positive response.