#Multimodal

AISGLang says it turned Qwen3.8-27B into a multimodal decision model that beat Pokémon FireRed's Elite Four and champion with sub-100 ms decisions from live game state. It introduces a native /v1/decisions endpoint for turning LLMs and VLMs into classification and scoring models. A /v1/systemone endpoint is also added so Jev-like open models can work with the TypeSafe SDK.

AIJerry Liu benchmarked gpt-6.1 sol on document OCR tasks and found a sizable increase in table parsing and reading order over gpt-6 sol from a week earlier. Its table parsing is similar to gpt-6 astra. He noted frontier models still cost roughly an order of magnitude more than cost-effective document parsing solutions, leaving room to improve the premium end above 1c per page.

AIHiggsfield AI introduced Dots, a joint OpenAI x Higgsfield product offering an always-on creative crew that keeps working while users are away. Users can check in by text, call, or email and pause the work at any time. The post says it is powered by GPT-6.1 Sol.

AIMicrosoft Research has introduced Quine, an early-stage research effort to build a multimodal world model of biology that connects insights across biological scales and modalities. The system is designed to help scientists computationally search a space far larger than intuition allows and prioritize hypotheses before lab testing. Experimental results are meant to feed back into the model and sharpen future research directions.

AIDeedy describes a video generation pipeline built around Opus 5.5 in Claude Code, routing image, video, audio, and TTS models through OpenRouter's single API key. The workflow adds reference-image consistency, animatics before full renders, a critic skill that screenshots and transcribes output for QA, and ffmpeg for most editing.

AIShanghai AI Laboratory's Intern-Decision family of 0.8B, 2B, and 4B multimodal models averages 79.38, 84.68, and 90.02 across seven decision benchmarks. Intern-Decision-4B scores 88.74, surpassing Jev while achieving better probability calibration. Reported mean latency is 33.98, 33.28, and 44.16 ms, versus 109.70 ms for Jev in the same local HF setup.

AIA user ran qwen3-vl:32b-instruct locally through Ollama on a 128GB computer to auto-tag 40,000 images in their Eagle app. The image library was collected over 10 years and is intended as a personal asset base for Claude Code-made knowledge videos. The user says the use case makes the 128GB memory purchase feel worthwhile.

AILlamaIndex argues that frontier vision-language models still fail on real forms such as W-2s, 1040s, W-9s, and scanned W-4s, because forms require detecting every field, preserving section hierarchy, linking values to their exact boxes, and reading handwriting and checkmarks. The company's blog post details these failure modes and presents a custom cookbook for LlamaParse as a cheaper way to handle such forms.

AIKling AI says its Kling 4.0 Flash is live now for Ultra Yearly subscribers, with the full Kling 4.0 coming this October. The update advertises up to 4K resolution, 10-bit HDR output, stereo audio, and native 30-second generation. It also adds Omni Reference supporting up to 15 multimodal references and multi-keyframe control with up to 10 keyframes.

AIJina-OCR-v1, a 3.4B-parameter MoE model that activates 570M parameters per token, converts entire document pages into structured Markdown at 2.57 pages per second. It scores 91.14 on OmniDocBench v1.6 and 83.4 on olmOCR-Bench, 7.4 points above DeepSeek-OCR on the latter, and delivers the highest throughput among 14 evaluated systems at concurrency 32. The model is released under CC BY-NC 4.0, so commercial use requires permission.

AIDeedy argues every open source repository should have a roughly seven-minute explainer video like the one made for SQLite, covering its purpose, a high-level code map, a query's path through the codebase, core abstractions, and a real execution trace including join-order query planning. He says the video was generated with Opus 5.5 and Gemini 3.8 TTS, and he expresses amazement at how coherent and capable the model is.

AIAnthropic's Felix Rieseberg had Opus 5.5 create a parody of the "nihilistic" penguin, generating all models and textures from scratch. The model did this by using Blender as a Python library.

AIAnthropic's Noah Zweben shared a claymation video showing Claude Code's /remote-control feature, now made with Opus 5.5 after an earlier Opus 4.6 version. The feature, which lets users control Claude Code remotely, is rolling out to Pro users at 10% and ramping, with Team and Enterprise support coming later.

AIMeituan's LongCat team has released LongCat-2.5-Preview, a natively multimodal model with 1.6T total parameters, about 48B active, and a 1M-token context window. The model is built for long-horizon tasks spanning terminals, browsers, GUIs, spreadsheets, and design tools. It is available now through an API on the LongCat platform and a chat interface.

AIRunway co-founder and Co-CEO Alex Germanidis explains why he sees world models as the endgame for AI video, with applications in robotics simulation and real-time generation. He also discusses how Sora pushed Runway into an intense competitive sprint and his vision of a fully neural operating system where interfaces are generated as pixels rather than HTML and code.

AIOdyssey introduced Agora-2, its next-generation multi-agent world model, which supports up to 20 humans and agents interacting in a shared environment simulated in real time. A multiplayer research preview is available to try now.

AIOdyssey's research preview of Agora-2 lets up to 4 humans join forces against 16 agents in an environment simulated entirely in real time by a world model. The post states there is no game engine underneath.

AIGoogle's Philipp Schmid announced Gemini 3.8 Flash, recommending Gemini for multimodal understanding. A quoted post by Spencer Schiff reported that frontier models struggled to match correct names to people in a drawing, offering a visual test for future models.

AIModelScope released Qwen-Image-2.1-Fun-Controlnet-Union, a single checkpoint adding eight structural controls, including Canny, Depth, Pose, and Scribble, plus inpainting to Qwen-Image 2.1. Control and inpainting share one branch with 16 injection points across every second Transformer block, keeping the base model frozen and requiring no checkpoint switching. It runs at guidance scale 1.0 with CFG-distilled sampling and prefix KV caching, and is available under the Qwen Research License with base Qwen-Image 2.1 weights required.

AITencent Hy Translation has launched, powered by Hy-MT2, supporting 33 languages and 5 Chinese minority languages and dialects. It offers voice and photo translation with full offline, on-device operation requiring no network, and is already live in 12 countries and regions.

AIMeta's Muse Realtime Avatar was rated ahead of two leading commercial avatar systems in live-call products, based on 2–3 minute conversations with matched avatar identities. Raters compared visual quality, sync, character consistency, and mannerisms, and Muse Realtime Avatar came out ahead on overall preference.

AIMeta's Muse Realtime Voice generates speech tokens encoding both content and prosody, and Muse Realtime Avatar consumes that shared stream to produce streaming video. Using a fixed-length history as motion context keeps computation bounded regardless of conversation length while synchronizing voice, lip motion, and expressions.

AIMeta unveiled Muse Realtime Avatar, a real-time embodiment technology that turns Muse Realtime Voice into expressive, interactive avatars. It enables new interactions in Muse, starting with real-time conversations.