Skip to content

Multimodal

Beyond text: vision, mixed image-text, audio and video input and output in models and products.

Latest picks

461–480 of 592

Apr 27Monday

Synced · WeChat

From 99 Lines of Frozen Code to Meshy AI’s 3D Momentum in the West

Meshy AI released Meshy 6 and claims over 60% share in developed Western markets. The post says it has 10M+ users, $40M+ ARR, and 100M+ AI-generated 3D models in three years. The key signal is workflow fit: 37Games reports 30–40% less base sculpting work.

Why it matters: Meshy 6 ships with the company claiming over 60% of the US and European 3D AI market and a 30% to 40% cut in base sculpting work at 37 Interactive.

Apr 25Saturday

Hacker News front page

Google Flow Music

Google Flow Music launched a web creation entry with six sections: songs, playlists, Spaces, videos, projects, and Turntable. The page says Producer creates full songs with Lyria 3, and AI music videos use Veo. Pricing, regions, model specs, and rights terms are not disclosed.

Why it matters: Google launched the web version of Flow Music, putting Lyria 3 song generation and Veo music videos behind one entry point, with the conversational Producer tool making full tracks with vocals.

TechCrunch · AI

ComfyUI hits $500M valuation as creators seek more control over AI-generated media

ComfyUI raised $30 million at a $500 million valuation. The RSS snippet says its tools give creators more control over AI image, video, and audio generation; the post does not disclose investors, round stage, pricing, or release timing. The real signal is workflow control, not another model vendor.

Why it matters: ComfyUI raised $30 million at a $500 million valuation, betting that creators want to wire up node-based image and video workflows themselves for finer control over the output.

Hacker News front page

TIPSv2: Advancing Vision-Language Pretraining with Enhanced Patch-Text Alignment

Google DeepMind released TIPSv2 with 3 pretraining changes for CVPR 2026. iBOT++ applies self-distillation to masked and visible patches, adding 14.1 mIoU on ADE150; Head-only EMA cuts training parameters by 42%. The key signal is visible-token supervision, not a larger teacher model.

Why it matters: Google DeepMind's TIPSv2 adds self-distillation losses on both masked and visible patches, gaining 14.1 mIoU in zero-shot segmentation and cutting 42% of training parameters with Head-only EMA.

The Verge · AI

How Project Maven taught the military to love AI

In the first 24 hours of the assault on Iran, the US military struck more than 1,000 targets, with targeting accelerated by AI systems including Maven Smart System. The snippet says this was nearly 2x the scale of Iraq's “shock and awe” attack over 20 years ago, and Katrina Manson's new book traces Project Maven from its 2017 start in computer vision for drone footage; the post does not disclose model details, later contractors, or current deployment scope.

Why it matters: Katrina Manson's book traces Maven from 2017 drone-video analysis to Maven Smart System, whose Iran strike support hit over 1,000 targets in 24 hours, about twice Shock and Awe.

Apr 24Friday

Synced · WeChat

Remember more, answer faster, use less: HERMES speeds real-time streaming video understanding by 10x

Fudan University, Shanghai Academy of AI for Science, and NUS proposed HERMES, a training-free framework that turns KV cache into hierarchical memory for streaming video understanding and cuts TTFT by up to 10x. The post lists three mechanisms: hierarchical cache management, cross-layer memory smoothing, and position re-indexing; it reports 68% fewer video tokens with comparable or better results, and Qwen2.5-VL-7B on StreamingBench rising from 73.31% to 79.44%. What matters for practitioners: it answers without external retrieval, with TTFT around 27/29/28 ms at 16/64/256 frames.

Why it matters: Fudan University and other teams propose HERMES, a training-free framework that organizes the KV cache into three memory layers to speed up streaming video understanding, cutting first-token latency by up to 10x.

Synced · WeChat

After robots beat humans in marathon times: hardware nears its limit, intelligence becomes the second half

Honor's humanoid robot Lightning ran 50:26 at the 2026 Beijing Yizhuang half marathon, faster than the men's human world record of 57:20; the post also says Unitree H1 did a 1.9 km winding course in 4:13. The post cites nearly 200 embodied-AI financings and over RMB 30 billion in Q1 2026, plus Spirit AI's $455 million Pre-A on April 16. The real signal is capital shifting from robot hardware to model-centric 'brains.'

Why it matters: Honor's humanoid ran a half marathon in 50:26, faster than the human men's world record, while embodied AI drew nearly 200 funding deals worth over 30 billion yuan in Q1.

Xinzhiyuan · WeChat

Google's Vision Banana aims to unify vision tasks with a single pixel-generation interface

Google DeepMind and collaborators including Kaiming He introduced Vision Banana, claiming one pixel-generation interface can cover detection, segmentation, generation, and editing. The RSS snippet gives two head-to-head numbers versus Nano Banana Pro: 53.5% human win rate on GenAI-Bench and 47.8% on ImgEdit; it says only a small amount of reversible-format task data was mixed in, while data scale and full benchmark tables are not disclosed in the post.

Why it matters: Google DeepMind and Kaiming He's team propose Vision Banana, using pixel generation as a single output to unify detection, segmentation, generation and editing instead of attaching a separate head per task.

Hacker News front page

GPT-5.5: Mythos-Like Hacking, Open to All

XBOW says GPT-5.5 cut miss rate to 10% on its real-vulnerability benchmark, versus 40% for GPT-5 and 18% for Opus 4.6. It scored 97.5% on visual acuity and used about half the login iterations of the next-best model. The key point is black-box testing: GPT-5.5 without source beat GPT-5 with source.

Why it matters: Security firm XBOW tested GPT-5.5 against a real vulnerability set, and its black-box miss rate fell to 10%, below GPT-5 white-box at 40% and Opus 4.6 at 18%.

Apr 23Thursday

QbitAI · WeChat

Qwen3.6-27B open-weights, beats its 397B flagship predecessor on agentic coding

Qwen released Qwen3.6-27B and says it beats Qwen3.5-397B on 4 agentic coding benchmarks with about 1/15 the parameters. The post cites SkillsBench rising from 30.0 to 48.2, GPQA Diamond at 87.8, and AIME26 at 94.1; it uses a dense architecture, Thinking Preservation, and Gated DeltaNet, with weights on Hugging Face and ModelScope.

Why it matters: Qwen released Qwen3.6-27B, a 27B dense model it says beats its own 397B previous flagship on four agentic coding benchmarks.

Xinzhiyuan · WeChat

Tashi Zhihang raises $455.0 million in a Pre-A round, with Sequoia China and Hillhouse jointly leading

Tashi Zhihang said on April 16 it closed a $455.0 million Pre-A round led by Sequoia China, Hillhouse Ventures, and Meituan, which the post says set China records for embodied AI single-round and Pre-A financing. The post also says its AWE3.0 four-modal model lifted unseen-view task success by 3x and cut execution jitter by about 45%, and that its A1 robot set a Guinness record in sub-millimeter wire-harness assembly within one hour. What matters is whether model, data, and deployment keep reproducing; the post does not disclose valuation or deal terms.

Why it matters: Tashi Zhihang's $455m Pre-A, led by Sequoia China, Hillhouse and Meituan, sets a domestic record for a single embodied-AI round and comes with AWE3.0 and A1 deployment figures.

Hacker News front page

Website streamed live directly from a model

Flipbook generates an entire clickable website in real time with an image model, where each page is a pixel image and every click spawns a deeper image. The post says all on-screen text is drawn by the image model with no HTML or text overlays, and content comes from agentic web search plus model knowledge. The key point is the interaction model, not a standard generative UI; the live video stream remains an experimental, resource-heavy toggle.

Why it matters: Flipbook turns a webpage into a live-generated pixel image, where clicking anywhere generates a deeper new image and text is drawn by the image model itself.

Apr 22Wednesday

Hacker News front page

Qwen3.6-27B: Flagship-Level Coding in a 27B Dense Model

Qwen released the open-weight 27B dense model Qwen3.6-27B and made it available in Qwen Studio. It scores 77.2 on SWE-bench Verified vs. 76.2 for Qwen3.5-397B-A17B, and 59.3 on Terminal-Bench 2.0 under a 256K context and 3-hour timeout. The real takeaway is deployment: this is not a larger MoE, but a denser 27B model with stronger coding results.

Why it matters: Qwen open-sourced Qwen3.6-27B, a 27B dense coding model that scores 77.2 on SWE-bench Verified, beating the previous 397B-parameter MoE flagship, making the dense-versus-MoE comparison the draw.

QbitAI · WeChat

SenseAuto's Sage with 3B active params claims to beat GPT-5.4 and Opus 4.6 in cars

SenseAuto released Sage, an in-car multimodal edge model with 32B total params and 3B active params, and says it scored 94% on PinchBench, above Claude Opus 4.6 at 93.3% and GPT-5.4 at 90.5%. The post says Sage runs on Nvidia OrinX with about 0.5s TTFT, 0.03s TPOT, and 80 tok/s throughput; its SCOUT training method cuts GPU hours by about 60%, and ERL raises complex-task completion by 20%. The key point is not the headline race but whether a 3B-active model can sustain multi-step tool use on device.

Why it matters: SenseAuto's Sage is an on-car multimodal model with 32B total parameters that activates 3B per pass, scoring 94% on PinchBench ahead of Claude Opus 4.6 and GPT-5.4, with OrinX latency and throughput numbers.

Latent Space

OpenAI launches GPT-Image-2

OpenAI shipped GPT-Image-2 in ChatGPT, Codex, and the API. It has thinking and non-thinking variants, with stronger text, layout, editing, multilingual output, and QR codes. Arena ranks it first on 3 Image Arena boards, with 1512 Elo in text-to-image and a +242 lead.

Why it matters: OpenAI put GPT-Image-2 into ChatGPT, Codex and the API, with better text rendering, layout editing and multilingual output for slides, infographics and UI mockups.

X · @dotey

OpenAI launches ChatGPT Images 2.0, available to all ChatGPT and Codex users starting today

OpenAI made ChatGPT Images 2.0 available today to all ChatGPT and Codex users, and also opened the gpt-image-2 API. The RSS snippet says it supports up to 2K output, aspect ratios from 3:1 to 1:3, and more reliable non-English text rendering. In thinking mode, it can search the web, generate multiple styles, and self-check outputs; that tier is limited to Plus, Pro, and Business, with Enterprise not yet available.

Why it matters: ChatGPT image generation reaches 2.0 with free-tier access, steadier complex layouts and multilingual text, up to 2K output and aspect ratios from 3:1 to 1:3.

X · @OpenAI

Introducing ChatGPT Images 2.0

OpenAI introduced ChatGPT Images 2.0 as an image model for complex visual tasks and directly usable visuals. The RSS snippet cites sharper editing, richer layouts, and “thinking-level intelligence,” but the post does not disclose model size, pricing, latency, or rollout scope.

Why it matters: OpenAI's ChatGPT Images 2.0 announcement claims finer edits, richer layouts and thinking-level intelligence, but gives no model size, pricing or rollout details.

Bloomberg Technology

OpenAI unveils new image model that is better at charts and diagrams

OpenAI released an update to its image generation software to produce more accurate, complex charts and scientific diagrams. The RSS snippet does not disclose the model name, launch timing, pricing, benchmarks, or technical method. The real signal is a push into professional use cases, not generic image quality.

Why it matters: OpenAI's image model update targets accuracy on complex charts and scientific diagrams rather than general image quality.

The Verge · AI

OpenAI’s updated image generator can now pull information from the web

OpenAI said ChatGPT Images 2.0 can pull information from the web when a thinking model is selected, helping generate multiple images from one prompt. It runs on GPT Image 2 and is available to ChatGPT Plus, Pro, Business, and Enterprise users; the post does not disclose rollout timing, usage limits, or pricing changes. The key shift is web-grounded multi-image generation, not just image quality.

Why it matters: ChatGPT Images 2.0 can now search the web before generating a set of images, available to Plus, Pro, Business and Enterprise users.

TechCrunch · AI

Report says Clarifai deleted 3 million photos OkCupid provided to train facial recognition AI

Clarifai deleted 3 million photos from OkCupid after an FTC settlement, and the images had been used to train facial recognition AI. The RSS snippet says the data sharing request dates to 2014 and OkCupid executives had invested in Clarifai. The post does not disclose the settlement terms, deletion verification, or model impact.

Why it matters: Clarifai's FTC settlement forced it to delete 3 million OkCupid user photos and the facial recognition model trained on them, data OkCupid handed over in 2014 while an executive was investing in Clarifai.