Skip to content

Multimodal

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

Latest picks

521–540 of 592

Apr 7Tuesday

Latent Space

[AINews] Gemma 4 crosses 2 million downloads

Google’s Gemma 4 reached about 2 million downloads in its first week. The post compares that with Gemma 3 at 6.7 million over the past year, Gemma 2 at 1.4 million since June 2024, and Qwen 3.5 at about 27 million in roughly 1.5 months. The signal for practitioners is local deployment: one iPhone 17 Pro demo ran Gemma 4 E2B at about 40 tok/s via MLX, with support across Hugging Face, vLLM, llama.cpp, Ollama, and NVIDIA.

Why it matters: Gemma 4 hit about 2 million downloads in its first week, against 6.7 million for Gemma 3 over a year and about 27 million for Qwen 3.5 in six weeks.

Apr 3Friday

X · @op7418

Alibaba released the Qwen 3.6 Plus model

Alibaba released Qwen 3.6 Plus with a 1M context window, 64K input, and nearly 991K max output. The RSS snippet says it improves over Qwen 3.5 on agents, coding, image, and document understanding, priced at RMB 2 per 1M input tokens and RMB 12 per 1M output tokens; benchmark scores and test conditions are not disclosed.

Why it matters: Alibaba put Qwen 3.6 Plus on Bailian, raising context from 256K to 1 million tokens, with stronger agent and coding ability than 3.5 and better image and document understanding.

X · @op7418

Google releases Gemma 4 for on-device use under Apache 2.0

Google released Gemma 4 in four variants—E2B, E4B, 26B MoE, and 31B Dense—targeting phones, edge devices, and up to single-H100 workstations. The RSS snippet says the 26B MoE activates 3.8B parameters and adds native function calling, JSON output, multimodal I/O, speech-to-text, and Apache 2.0 licensing; the post does not disclose benchmarks, context length, or rollout details.

Why it matters: Google released Gemma 4 in four sizes under Apache 2.0, with a 26B MoE activating 3.8B parameters per pass and native function calling, JSON output plus image, video and audio handling.

X · @dotey

LatePost on DeepSeek before V4: traits, organization, and Liang Wenfeng's goals

LatePost says DeepSeek has confirmed 4 core departures, and V4's large model slipped from around Lunar New Year to April; the report says it will likely remain open source. The snippet cites 2x-3x recruiting offers, some 8-digit packages, a 100-plus research team, and a shift from CUDA/Triton to TileLang for domestic GPU adaptation. The real signal is strategy: DeepSeek had spent less on agents and coding, but now names an agent product role; the post does not disclose V4's size, price, or benchmarks.

Why it matters: LatePost traced the departures of four core DeepSeek members and the team's route choices before V4, giving context on the company's current organization.

X · @dotey

Google releases the Gemma 4 open model family under Apache 2.0

Google released the Gemma 4 family and switched the full line to Apache 2.0. The post says it includes 31B Dense, 26B MoE, E4B, and E2B; 31B and 26B support 256K context, and 31B fits on one 80GB H100. The key change is distribution terms: fewer limits on commercial use, modification, and redistribution, plus native function calling and structured JSON for agent workflows.

Why it matters: Google moved the whole Gemma 4 line to Apache 2.0, removing commercial, modification and distribution limits that its own license had left gray.

Google DeepMind

Google DeepMind releases the Gemma 4 open model family

Google DeepMind released Gemma 4, which it calls its most intelligent open model yet, aimed at advanced reasoning and agentic workflows under an Apache 2.0 license. The family comes in four sizes: E2B, E4B, 26B MoE and 31B Dense. The 31B ranks 3rd among open models on the Arena AI text leaderboard, and the 26B ranks 6th.

Why it matters: Gemma 4 ships under Apache 2.0 in four sizes, with 31B ranked 3rd and 26B 6th among open models on the Arena AI text board, a reference for deployment and fine-tuning choices.

Apr 1Wednesday

MIT Technology Review · AI

The gig workers who are training humanoid robots at home

Micro1 hires thousands of contractors across 50+ countries to film chores at home with iPhones and sell that real-world data to humanoid robotics companies. The piece cites $15/hour pay for one worker, says robotics firms spend over $100 million a year on such data, and notes $6 billion+ went into humanoids in 2025. The real issue is data governance: workers know the footage trains robots, but the post shows they often do not know how it is stored, shared, or deleted.

Why it matters: Micro1 pays gig workers in more than 50 countries to film household chores on iPhones and sells the motion data to humanoid robot companies; a Nigerian medical student earns $15 an hour.

TheValley101 (硅谷101)

E231 | From B2B to A2A: What Agent Infrastructure Could Do for a One-Person Global Business

Alibaba International president Zhang Kuo said procurement agent product Accio reached 10 million MAU in March and is still growing quickly month over month. The interview’s clearest metric: AI cuts procurement communication time to one-fifth, from about one week to one day, by chaining research, design-pack generation, cross-language communication, and supplier screening into an agent workflow. The real point is A2A: the post frames it as agents restructuring buyer, seller, and platform flows, not just a better chat box.

Why it matters: Alibaba International president Zhang Kuo says the sourcing AI product Accio cut cross-border procurement communication from about a week to one day, chaining market research, design drafts, cross-language communication and supplier screening into one workflow.

Mar 24Tuesday

Mistral AI

Mistral AI releases Voxtral TTS, a 4B-parameter speech model

Mistral AI released Voxtral TTS, its first text-to-speech model. It has 4B parameters and supports nine languages: English, French, German, Spanish, Dutch, Portuguese, Italian, Hindi and Arabic. It handles emotional expression and zero-shot cross-lingual voice adaptation.

Why it matters: A 4B open TTS model covering 9 languages, with a head-to-head against ElevenLabs as a reference point.

Mar 18Wednesday

MIT Technology Review · AI

The Pentagon plans to let AI companies train models on classified data, defense official says

The Pentagon is discussing secure facilities where AI firms can train military-specific models on classified data. The post says training would follow tests on nonclassified data; the DoD keeps data ownership, and company staff would access it only rarely with clearance. The key issue is leakage: one shared model may resurface classified information across groups with different access levels.

Why it matters: The Pentagon plans secure environments for OpenAI, xAI and others to train military models on classified data owned by the Defense Department, with company staff access limited to rare cases.

Mar 17Tuesday

OpenAI News

Introducing GPT-5.4 mini and nano

OpenAI released GPT-5.4 mini and nano on March 17, 2026 for coding and subagents; mini runs over 2x faster than GPT-5 mini. In the API, mini has a 400k context window and costs $0.75/$4.50 per 1M input/output tokens, while nano is API-only at $0.20/$1.25. The key signal is performance per latency: mini scores 54.4% on SWE-Bench Pro versus GPT-5.4 at 57.7%.

Why it matters: OpenAI released GPT-5.4 mini and nano for high-throughput, low-latency work, with mini beating the prior generation on coding, reasoning, vision and tool use at more than double the speed.

Mistral AI

Mistral releases Mistral Small 4, unifying reasoning, multimodal and coding

Mistral AI released Mistral Small 4, the first Mistral model to unify Magistral reasoning, Pixtral multimodal and Devstral coding-agent abilities in a single model. It ships under the Apache 2.0 license.

Why it matters: Merging reasoning, multimodal and coding into one open model, with latency and throughput numbers against Mistral Small 3.

MIT Technology Review · AI

Where OpenAI’s technology could show up in Iran

Just over two weeks after OpenAI’s classified-use deal with the Pentagon, MIT Technology Review outlined three places its tech could surface in Iran-related conflict. The post names target prioritization, Anduril counter-drone analysis, and GenAI.mil back-office use; it does not disclose when classified integration will finish or confirm deployment in Iran.

Why it matters: Two weeks after OpenAI signed a classified deal with the Pentagon, MIT Technology Review lays out three scenarios where its tech could be used, including target ranking and strike decisions.

Mar 13Friday

MIT Technology Review · AI

A defense official reveals how AI chatbots could be used for targeting decisions

A US defense official said the Pentagon can feed target lists into generative AI, have the model rank them using factors like aircraft location, and send strike recommendations for human review. The post says this chatbot layer may sit on top of Maven to speed search and analysis, but it does not disclose the speed gain, and the official did not confirm current operational use. The key issue is verification: chat outputs are easier to use than Maven’s map UI but harder to check.

Why it matters: A US defense official describes a workflow where a chat model ranks target lists and humans review, possibly layered on the existing Maven system.

Mar 9Monday

MIT Technology Review · AI

How AI Is Turning the Iran Conflict Into Theater

The author reviewed more than a dozen Iran-war dashboards in one week and argues they turn satellite data, ship tracking, AI summaries, and betting links into a real-time war spectator interface. The post cites a dashboard built by two Andreessen Horowitz staffers that pulls in Kalshi bets, while Craig Silverman has logged 20 similar dashboards. The point to watch is information quality: the piece cites Financial Times reporting on AI-generated satellite images spreading online, while these dashboards lack the human vetting and historical context used by intelligence agencies.

Why it matters: A look at how dashboards built by a16z staffers and others turn satellite imagery, ship tracking and AI summaries into a live spectator view of the Iran conflict.

Mar 5Thursday

36Kr (direct RSS)

Embodied AI company Pascini raises over RMB 1 billion in Series B, valuation tops RMB 10 billion

Pascini said it closed a Series B round of over RMB 1 billion, bringing its valuation above RMB 10 billion. Lead investors include Huangpujiang Capital, Kaitai Capital, and Xinan Capital; the post says Pascini will use 10-billion-scale real-world multimodal data to train its VTLA model, but does not disclose model details.

Why it matters: A Chinese embodied-AI startup raising over 1 billion yuan at a 10 billion yuan valuation shows how much money is flowing into robot foundation models.

Feb 28Saturday

36Kr (direct RSS)

Qwen plans AI glasses, earbuds, and rings as tech giants race for a new AI entry point

A report says Alibaba's Qwen plans AI glasses, earbuds, and rings for a global launch in 2026; the glasses are slated for MWC 2026, with reservations opening on March 2. The post adds that Qwen app functions like food delivery and ride hailing will move to these devices, and cites Qwen3.5-Plus with 60% lower memory use, up to 19x inference throughput, and RMB 0.8 per million tokens. The real point is distribution: if the hardware connects Alipay, Amap, and Taobao, Alibaba is chasing the consumer AI entry layer, not just device sales.

Why it matters: Alibaba plans Qwen AI glasses, earbuds and a ring for 2026, moving food delivery and ride hailing from its app onto wearables, as model apps push past the phone.

Feb 27Friday

36Kr (direct RSS)

Embodied AI startup Zhongke Diwuji, which supplies the "brain" for Unitree, raised hundreds of millions of yuan

Zhongke Diwuji completed Pre-A and Pre-A+ rounds worth hundreds of millions of yuan within one month, and became a Unitree core ecosystem partner in Jan 2026. Since 2025, it has supplied the "brain" model for Unitree robots; the company says its FAM models use secondary pretraining and heatmap alignment to learn new tasks from 3-5 real-robot demos, with 97% success on basic tasks. The signal to watch is commercialization: it is moving from POC to power inspection, industrial handling, and retail deployments, charging robot OEMs per-device license.

Why it matters: Zhongke Diji raised two rounds worth hundreds of millions of yuan in a month and supplies brain models for Unitree robots in power inspection and industrial hauling.

Feb 26Thursday

New York Times Chinese

Where Is the U.S. Losing to China in AI?

The piece argues China has embedded AI into manufacturing, with 30,000+ smart factories, and over half of all industrial robots installed globally in 2024 going to Chinese plants. It cites shop-floor data: Zeekr's Ningbo plant uses 800+ robots, Xiaomi says its Beijing factory produces one car every 76 seconds, while only 18% of U.S. manufacturers report a formal AI strategy and two-thirds struggle to scale pilots. The real point is not frontier models but AI deployment in factory automation, scheduling, and inspection.

Why it matters: China has built over 30,000 smart factories, and more than half of 2024's new industrial robots went to Chinese plants, with Zeekr, Xiaomi and Midea showing AI on the line.

Feb 14Saturday

MIT Technology Review · AI

ALS stole this musician’s voice. AI let him sing again.

Patrick Darling, 32, returned to the stage on February 11 in London after two years without singing, using an AI voice clone rebuilt from old recordings. The post says speech cloning typically needs about 10 minutes of clean audio; his singing clone was built from noisy phone clips and kitchen recordings, then refined with Eleven Music over about six weeks. The practical signal is access, not sentiment: ElevenLabs offers the tools free to people who lost their voices to ALS and similar conditions, but the post does not disclose model details.

Why it matters: A 32-year-old musician with ALS sang with his band again using an AI voice clone built from noisy phone video and kitchen recordings.