@GoogleDeepMind
DeepMind introduces AlphaGenome 2
From tissue averages to individual cell types: researchers can specify a cell type or state when predicting how DNA variants affect gene regulation.
The model predicts how the same DNA variant acts in different cell types. On the TenK10K immune-cell dataset, DeepMind reported identifying 30% of causal variants, up from 16%. Researchers can supply single-cell data to study new cell states without retraining the model. The API is open for noncommercial research, with a private commercial preview; the model is not validated or approved for clinical diagnosis.
@runwayml
Runway opens the weights for its Praxis-1 robot model
The 27B model learns from video pretraining and predicts both frames and robot actions. Its weights, inference code and fine-tuning code are available under Apache 2.0. Runway reported a 46.2% success rate across 1,200 RoboLab-120 episodes, a result specific to that benchmark.
@sundarpichai
Gemini Decisions API scores up to 1,000 questions per request
Gemini Decision 1 scores predefined options directly. A request can evaluate up to 1,000 independent questions against the same text, images, audio, video or PDFs, using yes/no checks, choices and rubric scores. Google lists pricing at $0.10 per million input tokens and no output charge.
@perplexity_ai
Perplexity red-teams its agent sandbox with 12 models
The team asked 12 models to escape a guest VM and, separately, to reach a URL blocked by network policy. No run escaped to the host; agents found two ways around the policy in the pre-remediation configuration with partial network access.
@OpenAI · via @jeffreygwang
OpenAI studies how training changes chain-of-thought monitoring
The researchers used reasoning without chain-of-thought tokens and the ability to manipulate those tokens as proxy measures. Lower base-model perplexity correlated with stronger reasoning without a chain of thought. Reinforcement learning increased controllability in the experiments, and model-generated training data could pass that ability on through supervised fine-tuning.
@drjimfan
GR00T 2 demo shows dexterous hands running a lab task
NVIDIA researcher Jim Fan shared a video of real robot hands performing a chemistry experiment and cleaning up afterward. He described the footage as autonomous, uncut and shown at normal speed. It demonstrates one task and does not establish a success rate across tasks.
Qwen Code 1.0 enters preview
The 1.0 preview brings together Qoder and Qwen Code's coding capabilities. It supports resuming sessions and non-interactive execution, and ships native packages without requiring a separate Node.js installation. Interfaces and configuration may still change.
n8n Labs previews its open-source personal agent Nod
The research preview can run on your computer with a hosted or local model. The code uses Apache 2.0. Example tasks include preparing for calls, collecting invoices and managing subscriptions.
LlamaIndex details its LlamaParse penetration test
Over four days, agents with a standard account probed parsing, extraction, indexing, deployed agents and MCP endpoints. LlamaIndex noted that repeat runs could take different paths even with the same prompt.
Hugging Face shares an agent-trace fine-tuning example
The example uses coding-agent trajectories as supervision in TRL and shows how to run the resulting model in a coding agent.
v0 adds support for ChatGPT Business subscriptions
v0 says it now accepts ChatGPT Business subscriptions alongside the existing Plus and Pro options.
Stanford shares materials for its agent engineering course
CS329Z has begun releasing lectures and assignments. The first assignment asks students to build a command-line coding agent from scratch.
Computer-use agent lectures survey research and evaluation
Jing Yu Koh shared lecture materials covering the history of computer-use agents, related benchmarks and research.
Vercel offers a limited-time trial of Beam
Guillermo Rauch says Reflection's open-weight Beam model is free on AI Gateway until November 12, with zero data retention. He also shared fx, a native coding agent written in Zig.
vLLM compares inference throughput on Rubin and GB300
In a specified AgentX setup running MiniMax-M3.1-Flash, vLLM reported 10.2× the total throughput per chip on Vera Rubin NVL72 versus GB300 NVL72, including cached tokens, at a P90 rate of 546 tokens per second per user.
vLLM separates the multimodal encoder into its own stage
The EPD pipeline splits encoding, prefill and decode, with remote embedding transfer and caching across encoders. The team reported lower time to first token and higher per-GPU throughput in tests that counted the encoder GPUs.
SGLang tunes Muse Spark Open inference on TPU
The SGLang team says the model runs on a single TPU v7 host. In the AgentX workload, a custom kernel and DFlash2 delivered about 3× faster decoding at small batch sizes.
vLLM-Omni adds day-one support for Praxis-1
The inference stack supports Runway's robot action model, taking camera observations as inputs and returning actions.
Tencent Hunyuan Hy4 adds hosted access
Tencent confirmed Hy4 availability on OpenRouter and the Atlas Cloud API. Inferact also announced a partnership with Tencent to host the model.
MiniMax reports adoption figures for H3
About 100 days after the H3 weights opened, a MiniMax team member reported 1,472 derivative models and more than 148 million Hugging Face downloads.
Transformers maintainers preview decision-model batching
Maintainers previewed batching for decision-model inference and reported roughly 3–4× speedups in early testing. The work is still in development.
vLLM passes 4,000 GitHub contributors
The project reported a community milestone of more than 4,000 contributors on GitHub.
Adobe launches Firefly Design in beta
The workspace combines vector, raster, motion, 3D and print work, from initial concepts to delivery.
Adobe previews local MCP integrations for Creative Cloud
At MAX, Adobe previewed local MCP integrations for Photoshop, Illustrator, Premiere and After Effects so outside agents can call into the tools.
Recraft updates vector generation and conversion
Vectorizer Pro converts user images to vectors. Recraft says generated and converted artwork now has smoother curves and cleaner corners, and the update is available on all plans.
Luma adds character replacement to existing shots
Character Swap & Insert can replace people or vehicles and add or remove objects in an existing video shot.
Runway adds 4K Seedance 2.5
Runway announced 4K video generation with Seedance 2.5 on its platform.
Runway demonstrates a creative workflow inside Claude
In Runway's demo, Fable 5.2 coordinated a workflow from concepts in Runway to an electronic flute prototype in Blender, then back to Runway for textures and renders.
PixVerse adds Seedream 5.5
PixVerse published a side-by-side comparison with GPT Image 2.5 covering composition, style, typography and detail. The platform selected the examples.
Firefly assistant beta can remember preferences across chats
Adobe demonstrated how the assistant can carry creative preferences across chats. Its trial announcement did not spell out the full availability.
Vidu showcases Q4 Preview for iterating on ads
Vidu used ad concepts to demonstrate Q4 Preview and reported faster generation and better prompt adherence. The announcement did not specify standardized test conditions.
fal demonstrates image and video editing inside ChatGPT
The fal plugin demo shows image and video generation and editing in a chat. The team used GPT Image 2.6 to generate keyframes.
ElevenAgents adds NVIDIA Nemotron models
NVIDIA and ElevenLabs expanded their collaboration, bringing Nemotron to ElevenAgents as another model option for enterprise voice agents.
Fish Audio Live 3 supports full-duplex speech
According to the vLLM project, Live 3 can continue listening while speaking. The vLLM team expressed interest in supporting a future open release.
Robots rehearse in simulation after one demonstration
The WICL paper stores simulator feedback in context without updating weights. The authors report an average success rate of 83.4% across clean and randomized settings for 50 RoboTwin 2.0 tasks.
Sentence Transformers 6.3 trains for vector compression
The release supports token-pooling-aware training. In one maintainer-run NanoBEIR experiment, retaining a quarter of the document vectors preserved 97.8% of the unpooled score, versus 94.6% with standard training.
Tree-Ising explores bit-level weight repair
The paper frames direct language-model weight repair as sparse discrete optimization to select which weight bits to change.
Community contributors release 47 lemma-level RL tasks
Following the Hugging Face release of a math reinforcement learning environment, a community contributor broke selected problems into 47 smaller lemma-level tasks.
Developer uses a 0.8B model to OCR 14,483 pages
A developer reported processing 14,483 pages of industry documents with LightOnOCR-3-0.8B. The post describes a workflow rather than a comparative accuracy benchmark.
e-means cuts iterations in large embedding clusters
On several embedding datasets, the authors report that e-means matched or beat 25 epochs of Lloyd k-means in about 1.4–2.6 epochs. The gain depends on the dataset and implementation.
Import AI reports a machine-assisted quantum proof
The newsletter reports that AI science company PSI used an internal model to obtain a proof of the Aaronson–Ambainis quantum conjecture and checked it in Lean.
Mistral live-transcribes conference talks
At AI Engineer Paris, Mistral used Voxtral Transcribe 3 on main-stage talks, supplying speaker names and technical terms as context.
Cerebras demos a shopping workflow with Kimi
Cerebras demonstrated Kimi K3.1 Pro using a single prompt to find a monitor and proceed to checkout. It was a product demo.
World Labs demo turns one image into an explorable space
A developer shared an example of building an explorable 3D scene from a still image, using Gaussian splatting to represent the completed space.
Interior AI demos room edits from video and floor plans
Its developer demonstrated removing or moving furniture and recoloring walls, floors and ceilings from uploaded video, photos or a floor plan.
A speed-sketch app demos Gemini's closed-set decisions
The AI Studio demo sends a sketch to Gemini Decision 1 and asks the model to identify the most likely object from a shortlist.
Creator demos a SQL ray tracer written with Carbon
A creator shared an experimental 3D ray tracer written in SQL with Gemini 4 Carbon.
Creator records Fable 5.2 painting stroke by stroke
The creator says the model took about 94 minutes and 11,066 strokes to paint a digital image in a browser tool, and generated music for the time lapse.
Designing agent interfaces for human review
Using Claude Artifacts as an example, a developer argues that interfaces should help people locate changes, review results and make adjustments after an agent acts.
Ask long-running agents for early intermediate outputs
A creator recommends requesting intermediate outputs so people can steer a long task while it is still underway.