Kimi K3.1 Pro opens its weights

OpenAI's developer team announced GPT-Image-2.6 for Codex and the API. It reports 50% faster generation for Flare and Sunburst, with Sunburst preserving details more consistently across edits at medium or high quality.

vLLM announced day-one serving support for Kimi K3.1 Pro and its million-token context. The cited 65.7% Terminal-Bench 4.0 score uses the Claude Code harness; it is a model benchmark, not a vLLM serving-throughput measurement.

SGLang says Kimi K3.1 Pro replaces dense MLA with DSA sparse attention. Its day-one figures include up to 33.7K tok/s per GPU on AgentX and GB300, and 450 tok/s for one user with DFlash2 and MegaKernel. These figures refer to different test configurations.

Ollama says Kimi K3.1 Pro is available on its Cloud, describing 2.8T total and 104B active parameters for coding, 3D creation and research. It advertises US hosting with zero data retention. Cloud availability does not establish equivalent local-device support.

According to @Kimi_Moonshot: Nebius Token Factory launches Moonshot's open-weight model on day one, focused on coding performance.

Tencent Hunyuan reposted Agent Arena's statement that Hy4 ranked 14th overall at that time and remained proprietary. The excerpt cuts off during the discussion of weights, so it does not confirm a release date for them.

According to @_lewtun: The 35B open agent base model provides reproducible checkpoints, 18K SFT samples, and a three-stage training recipe, supporting vLLM/SGLang.

Hugging Face reposted Niels Rogge's report of reproducing a paper with ML Intern, an open-source harness for AI research and model training. The excerpt does not identify the full paper or its results, so no specific metrics are inferred.

Stas Bekman shared DeepSpeed PR #8949, preserving gradients across request boundaries until the actual optimizer step. The author recommends deepspeed >= 0.19.9; the linked PR supplies the scope of the fix.

According to @woosuk_k: Alongside the Kimi K3.1 Pro release, the two sides are expanding collaboration to optimize serving and deployment of the model.

LlamaIndex reposted Jerry Liu's claim that OCRia-1 document-parsing throughput rose from 285 to about 2,800 tok/s on one B200. The roughly 10x figure is specific to the reported setup, not a universal speedup for every document workload.

LoCo identifies overthinking states in a model's own math trajectories, creates shorter verified continuations and trains local preferences with DPO. The authors report 41.7%–45.4% shorter math outputs and 23.4%–36.3% shorter code outputs across five math and three code benchmarks, with accuracy maintained or improved within those tests.

According to @runwayml: Developers can generate Runway video, image, and audio API integration code directly in Claude Code.

According to @synthesiaIO: Users can ideate, generate, and automate videos in ChatGPT, taking ideas to finished videos via a plugin.

According to @AravSrinivas: Uses the MotionSites skill to generate websites with WebGL shaders and scroll animations.

According to @amasad: An operations lead built an app with Replit that handles support ticket priority, context completion, and workflow routing.

Musk says Grok Bot can research purchases, bring back options, order items, track them and return unwanted goods. The post gives no availability scope, success rate or permission mechanism; this is a vendor capability claim.

Musk shared a user's account of having Grok Bot call an EV-charger installer for a quote and checking usage. The excerpt lacks the full metering figures, so it does not establish a call price.

Greg Brockman says OpenAI found, validated and prepared patches for nearly 1,000 Linux local privilege escalations at about $1,000 per vulnerability. He presents it as an emerging approach to software hardening, adding cost context to the same pipeline announcement.

Baoyu reposted huacnlee's experience that retaining target directories avoided clean rebuilds during AI development across multiple worktrees and reduced machine slowdowns. It is a personal setup report, not a universal disk-optimization result.

Baoyu describes a DeepSeek Harness plugin update for baoyu-design, adding a common entry for Slides, Design, Dashboard and Motion outputs. The project runs as an Agent Skill and produces self-contained HTML locally with tools including Cursor and Codex.

The open-source native macOS app Cicada offers live bilingual subtitles, interpreting and AI notes for calls and streams. Its repository describes use with Zoom, Tencent Meeting, Feishu and YouTube, with translated speech output through a virtual microphone.

Baoyu shares a Google DeepMind hiring note focused on AI Studio and the Gemini API. The short post lacks role requirements and application conditions, which require the original vacancy listing.

S D Ebrahimi reposted Kory Mathewson's call for BS, MS or PhD student researchers at Google DeepMind Montreal. The research description is truncated in the input, so no additional topic is inferred.

According to @gdb: The tweet says ChatGPT can help non-experts design prototype devices, expanding physical engineering applications.

Stas Bekman reposted Karthik’s argument that product companies should be able to train models for their own agent harnesses, as frontier labs do. The excerpt cuts off before the tool description and does not establish a new service launch.

Anthropic describes its biology team working with Claude to map the first 3D structure of a vault particle from bacterial life. The announcement presents the barrel-shaped protein complexes as a longstanding mystery; the supplied excerpt does not establish that Claude carried out the research independently.

Omar Khattab reposted an experiment reporting 65 wins, 34 losses and one draw in 100 head-to-head matches after editing 0.0002% of NanoJev's weights. The task name is truncated in the input; this is a specific experiment, not a general guarantee for weight editing.

Omar Khattab reposted Gabe Pereyra's comments on MIT research, with the excerpt mentioning GPT-6 Astra reading Qwen model weights. Full experimental details are absent; this does not establish reliable understanding or editing of arbitrary models.

Omar Khattab reposted researchers exploring weight-space methods for vision-language-model safety alignment. The excerpt cuts off at the test name and supplies no full results, so success is not established.

Omar Khattab reposted @jacsebl's account of using fable to handwrite weights for a 100K-parameter model that distinguishes real and fake legal citations. No accuracy or independent testing is supplied; the post does not establish a professional legal-verification product.

Omar Khattab thanks the Harvey team for supporting weight-space programming research and mentions its open contributions to related work. The input supplies no funding amount or partnership terms.

LeCun reposted Numerama's coverage of work with Randall Balestriero, referring to a paper posted October 29. The excerpt cuts off in its small-model discussion and does not support a complete claim about reasoning performance.

LeCun reposted Randall Balestriero’s notice of a talk on recent JEPA research at INFORMS in San Francisco. It is a conference announcement, without a full paper or new experimental data in the supplied excerpt.

LeCun reposted his view that LLMs accumulate substantial declarative knowledge, enabling them to answer many questions. The excerpt cuts off before the rest of the argument; no broader conclusion about reasoning is inferred.

Arvind Narayanan reposted Sayash Kapoor's small-scale tests suggesting simple prompts let Codex evade Pangram. This is a limited reported result, not evidence that every text can reliably bypass detection.

According to @Tim_Dettmers: The tweet discusses whether ML compilers and domain-specific languages remain necessary after kernel agent experiments.

Ethan Mollick highlights an AI-semiotics paper suggesting visual symbols affect models in ways their own chains of thought may not recognize, with reactions differing from human interpretations. This is his reading of the paper, not an independent replication.

teortaxesTex reposted Yandex's announcement of Yambda-5B for developing and evaluating recommendation models at scale. The input lacks full dataset composition and licensing details, which require the dataset documentation.

According to @_akhaliq: Input a scene description or sketch to generate a full-screen Three.js 3D world explorable in the browser.

According to @_akhaliq: Edit video with text or a single frame and rerender all shots, aimed at long-form multi-shot scenes.

_akhaliq reposted Deemos' HYPER3D Rigging announcement for humans, animals and fantasy creatures, with prompt-based editing. The excerpt cuts off in the detail description, so no specific finger-level capability is inferred.

Kling highlighted Ash Koosha's use of Kling 4.0, judging generated shots by whether they could belong in a feature film. This is a creative criterion and case study, not a standardized quality benchmark.

According to @ViduAI_official: Generates cinematic scenes such as sunset duels, bamboo forests, and snowfield solos via lighting, composition, and color prompts.

Tom reposted Unreal Engine's announcement of a 5.9 Preview due the following week, including improved AI workflows and web targeting. A preview announcement is not a stable release.

CoffeeVectors reposted Mirage's claim that Opus 5.5 or Astra can use Tesseract to create 3D outputs with path-traced rendering. “Professional-grade” is promotional wording, with no independent quality evaluation supplied.

Linus Ekenstam shared a scene made with Mirage's Tesseract, appearing as a game character talking to a fisherman. It demonstrates a creative workflow, without establishing general quality for dialogue, world generation or rendering.

CoffeeVectors reposted creator Kevin Parry's concern that one-click AI copying makes original authorship harder to identify. It is an industry observation, not a measured platform-wide trend.

CoffeeVectors reposted a creative tip: when AI characters sound unnatural, examine how dialogue is written and directed in the prompt. This is practice-based advice with no quantified performance promise.

歸藏 comments on Muse's performance in some realistic scenes and suggests free access helps its spread. The input is personal experience, without a full model identifier, availability scope or comparative benchmark.

Glif announced AI model Claudia for casting in ads and microdramas, claiming 44M views. Its site describes a chat-based creative agent combining image, video and audio generation; the view count is vendor-reported.

According to @sama: OpenAI opposes relegating humans to a role that AI thinks should be rewarded, emphasizing that humans should always control the future.

According to @fchollet: He says the future worth building keeps AI in human hands and uses it to improve the human condition.

According to @kaifulee: Enterprises can list AI pilots but struggle to state business results; the key is applying the strongest technology to the hardest problems.

Nathan Lambert says a public, free Interconnects dashboard organizes data on the open-model economy, while a new interpretation series will be paid. The dashboard and commentary have different access models.

According to @random_walker: AI capabilities and job definitions change every few months; avoid productivity illusions and threat-driven overreaction.

Musk reposted @poteto's AI Engineer talk preview, which mentions 30x more code. The excerpt cuts off in the quality discussion; it is not independently verified evidence of 30x productivity with unchanged quality.

According to @elonmusk: The tweet argues shopping is a weekly high-frequency pain point, closer to ordinary users than coding demos and well suited to agent automation.

Andrew Ng cites a LinkedIn study placing AI Engineer among the ten fastest-growing occupations in seven major economies during July–September. He also promotes AI Dev in New York on November 30–December 1. Top ten does not mean the single fastest-growing job.

According to @natolambert: Commentary notes Chinese labs consistently meet weight release timelines; future delays would be a signal.

CoffeeVectors reposted a developer's point that discussions of AI-code maintainability should not assume human-written code is automatically maintainable. The excerpt supplies a viewpoint, not a controlled comparison.

CoffeeVectors shared an attributed Hinton comment that AI is forcing a rethink of human intelligence. The rest of the quote is truncated; the topic is retained without completing the quotation or treating it as settled science.

歸藏 says App Intents lets developers choose app functions to expose to Siri. The input contains only this creator's account, without an Apple announcement, version or supported-device list; the specific new capabilities and rollout remain unverified.

橘子 relays figures of 320M Uber orders during Halloween week, with agents accounting for 500 delivery orders and four rides. No primary disclosure or counting method is supplied; the numbers remain unverified and do not establish industry-wide agent adoption.

_akhaliq reposted Voyager's promotion claiming lower prices than Higgsfield, with Seedance 2.5 cited at half the price. A separate overall “5x cheaper” slogan lacks comparison details; it is not treated as one consistent discount rate.