OpenAI introduces GPT‑6 Cyber

vLLM lists 280B total parameters and 14B active per token for the open base model, with 33.5T pretraining tokens. Its announcement distinguishes the 128K training context from support for up to 1M tokens in fine-tuning and inference.

SGLang announced day-one support for the model’s built-in ECHO draft mechanism and a context of up to 1M tokens. Its announcement lists separate GPU configurations for FP8 and BF16.

Soumith Chintala reports 68 hours of pretraining on GB300s and 17 million post-training rollouts over nine days. He says all three training stages used NVFP4 and the model is available to train and fine-tune in Tinker.

Tencent Hunyuan shared a third-party claim that Hy4 and Grok 4.7 at xhigh both scored 35 on the AA Intelligence Index. The comparison applies to that index and configuration.

Moonshot shared AMD’s announcement that Kimi K3.1 Pro runs on eight MI355X accelerators. AMD named ATOM and vLLM; the post provides no comparable throughput figures.

Adobe showcased Firefly Generate Music for tailoring soundtracks by mood, genre, energy, and tempo. Commercial use is subject to Adobe’s terms.

Runway added Google’s Gemini Omni 1.2 Flash for generating and editing clips up to 15 seconds. Runway also offers upscaling to 4K.

Meta says Muse Realtime Avatar gives its realtime voice experience an expressive on-screen character. The rollout covers the U.S. Muse app.

Luma says its MCP integration exposes preset workflows in Claude and ChatGPT, including generating angles from a product photo, swapping scenes, and adding an offer.

Photo Reframe fills in the area around an image so the same asset can fit a social post, an album cover, or an ad placement.

MiniCPM-OCR crops layout regions before reading the text. LlamaIndex reports that it beat the listed comparators on ParseBench’s content-faithfulness metric.

Pika describes H3 Flex as a model for fast iteration, with separate audio output and resolution up to 4K. The company claims a 15-second video can take as little as about seven seconds to generate and offers 15 free generations per day.

Replit and Sumitomo Mitsui Finance and Leasing announced a partnership combining leasing expertise with tools for building custom enterprise software.

Perplexity says its Agent API effort tiers are cheaper while matching or improving research quality in its testing. A new Ultra tier supports deeper searches.

Runway co-founder Cristóbal Valenzuela said new paid subscriptions in October were nearly four times the year-earlier number. The figure concerns new subscriptions rather than the total subscriber base.

Perplexity’s CEO said the company’s internal workspace had roughly 10,000 Teammates over the preceding 30 days, with about 1,000 active the previous day. The counts refer to agent instances.

System Duo uses fast decision models for triage, guard checks, and verification. The semantic router calls a reasoning model when a task needs it.

vLLM says Kimi K3.1 Pro runs on TPU v7x through vllm-torchtpu. The linked llm-d guide separately describes SGLang on TPUs and vLLM on GB300 GPUs.

C. Wolfe shared a talk on evaluating tasks without a simple ground truth. It covers human evaluation, quality guidelines, and how to apply those methods more broadly.

Warp founder Zach Lloyd discussed the team’s task-to-pull-request workflow, agent quality measures, and supporting infrastructure with swyx.

Neubig posted a recording of a live demonstration that starts an open-source software factory workflow from a single prompt.

Recraft says GPT Image 2.6 is available on its free plan, with Flare for quick drafts and Sunburst for consistent edits. Users can compare the output with Recraft V4.1.

Organizer swyx reported the opening of the event’s first edition in China, with participants from teams including Microsoft, MiniMax, and World Labs.

Citing Artificial Analysis, swyx argued that multiple independent groups should reproduce model evaluations to reduce bias from vendor-controlled setups.

Ethan Mollick observed that some groups see AI as a reason to redesign the business, while others expect existing processes to absorb it. Both may describe their work as AI transformation.

CoffeeVectors shared a creator’s experiment using Opus 5.5 to help build a Blender café and placing four Gaussian-splat characters in the same scene.

Baoyu shared uni-robot-agent, which uses a common harness to explore robot control with different general-purpose models. The short post does not provide enough results to rank them.

Linus Ekenstam described the ring’s interface: a single pending decision appears on the display and a gesture accepts or rejects it. The post is a personal account.

A researcher highlighted a large pretraining run from the open language model project Marin. The shared post does not include final metrics.

arXivBangers highlighted a mechanistic study of the circuitry behind in-context learning for few-shot classification in small transformers.