TwelveLabs raises $100M to bet video deserves its own foundation models

Good morning 👋 A $100M round just staked a claim that video needs its own foundation models, not a bolt-on to whatever text model you're already running. Amazon liked the bet enough to write a check into the same deal that made it the preferred cloud.

In today's issue:

• 🔭 TwelveLabs raises $100M betting video needs its own foundation models
• 🧠 Models & Releases: Gemma 4 goes voice-native on Reachy Mini, Nemotron and GPT-OSS land in GovCloud
• 🔬 Research Highlights: a scientific-judgment benchmark nobody's close to acing, and coding-agent leaderboards that don't replay clean
• 🚀 AI Startups: CarbonSix and Higharc close rounds alongside TwelveLabs
• 🏛️ Policy & Regulation: FTC takes aim at ideologically-tuned chatbots, Fable 5 comes back online
• 🛠️ Try This: give your agent stack a front door before the video agent shows up
• 🔗 Quick Links: Persistent buys Nagarro, Anthropic halves its price for California

🔭 THE ONE THING

🎬 TwelveLabs raises $100M to bet video deserves its own foundation models

TwelveLabs closed a $100M Series B co-led by NEA and NAVER Ventures, with Amazon among the backers, pushing total funding past $207M. The money locks in two bets: more work on Marengo and Pegasus, the video-native models behind what TwelveLabs calls its "Video Cognition System," and a multiyear AWS deal making it the preferred cloud, with new models launching first on Trainium. That's a real wager, that video deserves dedicated foundation models instead of becoming just another modality Gemini or GPT absorb by default. Amazon writing a check into the same round where it becomes the preferred infra partner tells you who actually wanted this deal closed. Whether Marengo and Pegasus can stay ahead of general multimodal models that get better at video every quarter is the open question, and $100M only buys TwelveLabs time to answer it.

🧠 MODELS & RELEASES

• 🎙️ Hugging Face / Cerebras put Gemma 4 31B behind Cerebras inference for real-time speech-to-speech, cascading it with Nvidia's Parakeet for hearing and Alibaba's Qwen3TTS for talking back. It's already running on the Reachy Mini robots, 9,000-plus of them out in the world. The code's open, so the pipeline is swappable end to end, not a locked demo. Hugging Face
• ☁️ AWS / NVIDIA / OpenAI Nemotron (Nano 9B, 12B, 30B, and Super 120B) and GPT-OSS (20B and 120B) landed on Bedrock in GovCloud. Open weights running in a compliance boundary that closed frontier APIs mostly can't touch. If your workload has to stay in GovCloud, that's a real menu now instead of a workaround. AWS

🔬 RESEARCH HIGHLIGHTS

• GeneBench-Pro OpenAI's new benchmark hands models 129 synthetic biology problems that would take a genomics postdoc 20 to 40 hours to work through: not running the stats, but deciding which analysis a messy dataset can actually support. Top score on the board is GPT-5.6 Sol Pro at 31.5%. Claude Opus 4.8 landed at 16%, Gemini 3.5 Flash at 8.1%. Builder read: scientific judgment is still the gap the benchmark race hasn't closed. Introducing GeneBench-Pro

• Human vs. LLM research taste A new paper checks where model-generated research ideas land against what actual papers propose, and the same pattern shows up across every LLM tested: models cluster around "bridge" ideas, stitching two known methods together, while human researchers spread across a much wider set of ways to frame a gap. The ideas aren't bad. They're narrow. arXiv:2607.01233

• Coding-agent benchmarks, audited GSO, SWE-Perf, and SWE-fficiency all score agents on real-repo performance optimization. Replay the reference patches and most don't even hold up: 39 of 102 GSO tasks pass cleanly, 11 of 140 on SWE-Perf, 411 of 498 on SWE-fficiency. Rankings flip depending on scoring rules (eight submissions disagree on 9 of 28 head-to-heads), and SWE-fficiency piles 58.5% to 82.8% of its score weight onto the ten hardest tasks. Investor read: the agents are probably better than these leaderboards suggest, since on replay-valid tasks, some submission matched or beat the reference patch 85.3% of the time. The instrument is what's broken. arXiv:2607.01211

🚀 AI STARTUPS

• TwelveLabs closed a $100M Series B, co-led by NEA and NAVER Ventures with Amazon in the round. The money's going into Marengo and Pegasus, the embedding and video-language models behind what they call a "Video Cognition System." Translation: making footage searchable the way text already is. If you've ever needed to find one clip inside a thousand hours of raw video, this is the bet that finally pays off. TwelveLabs
• CarbonSix raised a $40M Series A (DSC Investment and LB Investment co-leading) to push physical AI, robotic hands built for factory floors, out of demos and into actual production lines. It already has paying manufacturing clients, which is rarer than it should be in robotics. PR Newswire
• Higharc brought in $95M Series C led by Insight Partners, pushing total funding past $170M. The homebuilding AI platform is moving past design software into supply chain, partnering with US LBM to run material takeoffs at scale. Customers are reporting 25-50% faster time to community opening, per the company. PR Newswire

🏛️ POLICY & REGULATION

• FTC wants comment on a proposed policy statement that treats ideologically-tuned chatbots as a potential deceptive-practice violation under Section 5, the theory being that a bot skewing answers to hit an "ideological objective" breaks a consumer's reasonable expectation of accuracy. Published in the Federal Register July 1 on a 2-0 vote, comments open through July 31. Builder read: if your system prompt nudges tone on anything political, this is the test the FTC just wrote down. FTC press release

• Anthropic has Fable 5 and Mythos 5 back worldwide as of July 1, after the export controls the US government slapped on both models June 12 got lifted June 30. The trigger was an Amazon-discovered jailbreak that talked Fable 5 into writing exploit code; Anthropic couldn't verify user nationality fast enough to comply, so it pulled access entirely rather than risk it. Now Anthropic, Amazon, Microsoft, and Google are rolling out a shared framework for scoring how dangerous a given jailbreak actually is. If you shelved a Fable 5 integration over the outage, it's safe to pick back up. Anthropic

🛠️ TRY THIS

Give your multi-agent stack a front door before you add a video agent

TwelveLabs just raised $100M on the bet that video needs its own foundation model, not a vision plug-in bolted onto a text agent. If that's right, your agent roster is about to get one more member, and it needs somewhere to live.

1. Stand up a serverless gateway (API Gateway + Lambda works) that path-routes /agents/{agentId} to your existing agents plus a placeholder video agent.

2. Point a standard A2A client at the gateway's single domain and confirm discovery and routing work with zero client-side changes, per AWS's writeup.

3. Give the video agent its own memory store in AgentCore, tagged with tenant and content-type metadata so its context can't bleed into your text agent's retrieval, following this pattern.

4. Fire one query that should route to video and one that shouldn't. If the router guesses wrong, fix the path config before you plug in a real video model.

Prompt: Sketch a serverless A2A gateway with path-based routing for a text agent and a video-understanding agent, plus AgentCore-style memory scoped by tenant and content-type metadata.

Worth a look

• AWS A2A gateway Serverless, path-based routing behind one domain; existing A2A clients work unmodified. Read it
• AgentCore Memory metadata filtering Tag memory by tenant and content type across ingestion and retrieval, built for multi-agent setups. Read it

🔗 QUICK LINKS

• Persistent Systems is buying Nagarro in an all-cash deal, €81 a share, a 140% premium, to build a 46,000-person, $2.9B-run-rate shop chasing the AI-led modernization wave. Nagarro
• Anthropic and California cut Claude's price in half for every state agency, city, and county, plus free training. DMV and Medicaid teams are already running it. California
• AWS published a SageMaker walkthrough for BoltzGen, taking protein design from a quick validation run to production batch jobs. AWS
• Inscribe built an agentic reviewer on Bedrock that flags doctored financial documents in under 90 seconds. AWS
• HippoRAG is a neurobiology-inspired retrieval pattern on AWS: a Neptune graph and personalized PageRank standing in for flat vector search. AWS

That's the signal worth your time today.

Pradeep Perugu