@Alibaba_Qwen · Model release
Alibaba releases Qwen4-Max
The new flagship accepts native visual input and has a one-million-token context window.

Following last week's preview, Alibaba made Qwen4-Max available through the QwenCloud API and Token Plan. Third-party providers including OpenRouter also added the model. Alibaba describes a sparse mixture-of-experts architecture with 2.4 trillion total parameters and about 71 billion active per token.
On its own Work360 benchmark at the High reasoning setting, Alibaba reports that the model used 35% fewer tokens than Qwen3.8-Max and cut the average cost per task by about 60%.
@TencentHunyuan · Model release
Tencent launches low-cost Hy4 flash

Tencent listed API rates of $0.070 for input, $0.281 for output, and $0.003 for cache reads per million tokens. OpenRouter offered a temporary 50% discount. Tencent says its own blind test of everyday tasks put Hy4 flash on par with Claude Haiku 5.5 at high reasoning effort.
@huggingface · Inference
Hugging Face adds OpenRouter routing
OpenRouter can now select an inference provider for chat, vision understanding, and agent requests. Existing Hugging Face tokens and endpoints still work. Model pages show provider prices, speed, and availability. Image, video, and audio generation continue through the existing inference routes.
@OpenAI · Healthcare software
ChatGPT for Healthcare adds more than 20 plugins
The additions span clinical care, care coordination, and business operations.

OpenAI also updated clinical evidence search and computer-use agents that can act across applications. The company described work with Abridge, Ambience, Cigna Healthcare, and CVS Health on tasks such as handling referrals, moving documents between systems, and assembling information for care teams.
@OpenAI · Research access
OpenAI expands its free access program for researchers
Applications open November 23. OpenAI plans to award 50,000 grants to eligible applicants on a first-come, first-served basis. Each grant provides ChatGPT Plus to the recipient and up to two collaborators, for a maximum of 150,000 people in this round. The goal for the program as a whole is 200,000.
OpenAI also published an analysis of how more than 1,800 pilot participants used ChatGPT. Survey respondents estimated that without AI the same research work at the same quality would have taken nearly twice as long.
Craig describes Astra's role in a lattice construction
Physicist Nathaniel Craig revisited his October preprint on OpenAI's science blog. He proposed the strategy, and GPT-6 Astra helped turn it into a candidate lattice construction and develop the detailed proof. The work concerns a nonperturbative definition of the Standard Model of particle physics.
Perplexity launches Frontier Research

The initiative aims to automate scientific research by developing algorithms, models, and evaluations in the open. Perplexity lists multi-agent orchestration, reinforcement learning, continual learning, and systems engineering among its priorities.
OpenAI names four external assessment partners
Apollo Research, METR, Transluce, and Goodfire will examine internal models, systems, or evaluation materials under agreed scopes to test alignment and safeguards. Researcher Neel Nanda welcomed the pilots and argued that the access arrangements still fell short of employee-level access.
Sakana AI releases the JFinBench finance benchmark
JFinBench draws 321 multiple-choice, written, and chart-reading questions from the 2026 editions of six Japanese financial certification exams. The team evaluated 12 models and identified knowledge of Japanese financial institutions and practices as a common source of errors.
HubEgo-100K releases its first dataset batch
Hubdata published the first release on the Hugging Face Hub. Linus Ekenstam says it includes about 10,000 hours of first-person recordings of human activity, a resource for studying how robots can learn from human demonstrations.
Perplexity demos citations to individual filing lines

CEO Aravind Srinivas showed how Computer or Teammate can cite numbers in financial research. Readers can follow a figure back to the corresponding line item in the original filing and check the source.

@vercel · Model access
Vercel AI Gateway supports private models
Developers can register self-hosted, fine-tuned, or prerelease models and call them using a model ID with the private/ prefix. Private models share the gateway API, authentication, and request logs with catalog models.
Claude Code team explains how it verifies agent work
End-to-end tests exercise actual user paths, property tests check invariants, and differential tests compare implementations. The team also discusses how to detect false results and cases where a verifier can be bypassed.
Vercel distinguishes success from repeatability in agent evals
The post uses pass@k to measure whether at least one of several attempts succeeds, and pass^k to measure whether every attempt succeeds. It also describes an end-to-end evaluation setup using Harbor, Vercel Sandbox, and AI Gateway.
Claude Code desktop can test both iPhone Duo layouts

Anthropic announced a public beta that runs the iPhone Duo in an iOS simulator alongside the conversation. Claude can fold and unfold the simulated device to test the app in both states.
LM Studio adds remote control for Bionic agents

A Bionic agent can run on a laptop, a development machine, or a cloud VM while a user controls the session from another linked device. LM Studio says LM Link provides end-to-end encryption.
Replit's Microsoft Fabric integration is generally available

Teams can build apps with Replit Agent using their organization's OneLake data and deploy them to their Fabric tenant. The partner page lists support for Entra authentication and Purview governance.
Perplexity plans to move its Slack agent to Teammate
Srinivas says Perplexity's Computer agent in Slack will migrate to a persistent Teammate that can be assigned tasks. The announcement describes a planned migration.
Perplexity launches the Monitor API
Developers can set recurring checks for an event, topic, or page and receive updates via webhook. Srinivas announced a rate of $1 per 1,000 runs.
ParseBench ranks parsers by document needs

LlamaIndex demonstrated Metric Lab, which compares parsers on metrics relevant to a given job. Forms and tables may matter most for insurance claims, while charts may carry more weight for earnings presentations.
Cohere previews video search in Compass

Cohere is testing Video Intelligence in private preview. The feature is designed to make the contents of a video library searchable so teams can retrieve relevant moments.
Mollick points to three interface gaps in Dot
Ethan Mollick compared his experience using Dot, Muse, and Duet. He found that Dot's single conversation thread, fewer inline elements, and lack of a visible task list made it harder to use.

@_lewtun · Open-source experiment
More than 1,000 agents build a rule-based Python chatbot
Hugging Face had DeepSeek V4.1 Flash agents split the work of writing and testing PyBot. The released bot uses regular expressions, lookup tables, and branching rules to reply without loading model weights. The project reports a 26% win rate on the LLM-judged SmolTalkBench.
vLLM shares DeepSeek V4.1 Flash results on Vera Rubin

The project compared Vera Rubin NVL72 with GB300 NVL72 on SemiAnalysis's AgentX workload. Both used Dynamo and vLLM. The project reports that Vera Rubin was faster in that test.
Two DGX Stations serve GLM-5.3 NVFP4
NVIDIA published a guide that uses tensor parallelism to split the quantized weights across two GB300 GPUs and keep them in GPU memory. vLLM links the machines into one service with a single OpenAI-compatible endpoint.
MMFlash targets tokens that depend heavily on vision

The preprint finds that draft models are less likely to predict tokens that depend more on visual input. MMFlash gives those tokens more weight and uses sparse attention for long videos. The authors report up to a 3.58-fold lossless decoding speedup on the tested multimodal benchmarks.
ReDiverse aims to restore diversity after distillation

The authors link repetitive outputs from few-step generators to changes in activations during early denoising. One method fine-tunes a lightweight adapter. The other steers activations during inference using directions computed from the teacher model.
eDMD matches teacher distributions using final outputs

The proposed method trains one shared distribution estimator on the student's final samples and corrects the update signal at each denoising step. The aim is to reduce the influence of intermediate states on few-step distillation.
Raschka proposes a common API for model components

Sebastian Raschka argues that a building-block API for open-weight models would make it easier to reuse architecture components across models and to move experiments between PyTorch and direct C++ and CUDA implementations.

@synthesiaIO · Branded video
Synthesia introduces the Syren video engine
A team can supply brand guidelines and describe the desired video in chat. Synthesia positions Syren as a way to preserve brand appearance and pacing in launches, onboarding videos, and tutorials.
Higgsfield launches Short Drama Studio

Higgsfield says the tool turns a script into a multi-episode short drama while keeping the same cast across episodes. Its announcement lists 9:16, 16:9, and 4:3 formats.
Vidu demos character and outfit changes on live video

Vidu showcased the real-time mode of S2 Editing for applications that use a live camera feed. Developers can apply character, outfit, or style changes to the input stream.
Black Forest Labs launches a FLUX filmmaking challenge

Studios will write briefs for films made with FLUX. Black Forest Labs lists an online event from November 30 through December 13 and more than $25,000 in announced prizes. Partners include RTL Deutschland, Nura Studios, and Studio Babelsberg.
Runway and INK Studio partner on animated shorts
The two companies will use Runway's tools to make a series of animated short stories. The announcement gave no release date or division of production work.
A short-series creator starts with 480p clips
Creator @icreatelife described generating low-resolution footage in Seedance 2.5 with typically only one or two rerolls, then upscaling and editing in Premiere Pro. The workflow comes from the creator's own experience producing episodes daily.
Kling teases its Image 4.0 model

Kling shared a visual preview and said Image 4.0 was coming soon. The announcement gave no release date or API details.