This issue is sponsored by tastytrade.

Many brokerages were built for the passive investor. Tastytrade was built with active traders in mind, and their tools and features reflect that.

You can trade stocks, options, futures, and more on one platform, with low commissions including zero commission on stocks and crypto.

The real reason it stands out:

They give free educational courses if you want to sharpen your strategy.

Plus live Trade Desk support during trading hours from actual humans who know what they're talking about.

And yes, they've got receipts: awards from TradingView, StockBrokers.com, Investor's Business Daily, Bankrate, Investopedia, and more.

Join tastytrade and get up to three thousand dollars in commission rebates on your stock and ETF option trades for your first 30 days.

Mavericks don’t follow. They forge.

Visit tastytrade.com/Unleashed today.

Disclaimer: All applicable exchange, regulatory, and clearing fees still apply. Cryptocurrency services powered by zerohash. Zerohash receives a 50-75 bp markup/markdown of the executed order price of which tastytrade received 65%. Options involve risks and are not suitable for all investors. Learn more about tastytrade promo terms and conditions at tastytrade.com/double. Offer expires May 31st, 2026. tastytrade, Inc. is a registered broker-dealer and member of FINRA, NFA, and SIPC.

Mistral just introduced Mistral Medium 3.5, remote coding agents inside Mistral Vibe, and a new Work mode in Le Chat. The big idea is simple: instead of keeping coding agents trapped on your laptop, Mistral wants them running in the cloud, in parallel, continuing the job while you do something else. You can start these agents from the Mistral Vibe CLI or directly inside Le Chat, and they can keep working until they produce a branch, a draft PR, or a completed task.

The model behind this is Mistral Medium 3.5, a new 128B dense model with a 256k context window. Mistral describes it as a merged flagship model that combines instruction-following, reasoning, and coding into one set of weights. It is also released as open weights under a modified MIT license, and Mistral says it can be self-hosted on as few as four GPUs.

For coding, Mistral claims Medium 3.5 scores 77.6% on SWE-Bench Verified, which puts it in serious coding-agent territory. The model is also built for long-running tasks, structured outputs, and reliable tool calling, which matters because coding agents are only useful if they can actually work through messy real-world software tasks instead of just answering prompts.

The remote-agent feature is probably the most practical part of the announcement. Developers can inspect file diffs, tool calls, progress states, and questions while the agent works. Local CLI sessions can also be “teleported” into the cloud, carrying over session history, task state, and approvals. Vibe integrates with GitHub for code and PRs, Linear and Jira for issues, Sentry for incidents, and Slack or Teams for reporting.

Mistral is also expanding Le Chat with Work mode, a preview feature for more complex multi-step tasks. It can work across email, messages, calendars, the web, internal docs, and connected tools. Examples include preparing for meetings, researching topics, drafting reports, triaging inboxes, creating Jira issues, and sending Slack summaries. Sensitive actions still require explicit approval.

ElevenLabs officially introduced ElevenMusic, an AI-powered music discovery and creation platform built on top of its fully licensed music model. The platform is not just a song generator. It combines listening, remixing, original music creation, publishing, and monetization into one system.

Users can discover music, remix what they hear, or start from scratch with a lyric, melody, or mood. ElevenLabs says the system can help turn those ideas into full tracks by shaping structure and refining the sound.

At launch, ElevenMusic includes more than 4,000 independent and emerging artists, plus tracks from Eleven Album Vol. 2, including artists like Danger Twins and Justin Love. The idea is that fans do not just listen to the music; they can build on it, remix it, reinterpret it, and share new versions.

ElevenLabs is also pushing a creator-economy angle. Artists can publish original tracks or remixes, grow an audience, and earn based on listener engagement, eligibility thresholds, and platform revenue. The company says it has already paid out more than $11 million to creators through its voice library and wants to bring a similar model to music.

Figure says it has ramped production of its Figure 03 humanoid robot at BotQ, its high-volume manufacturing facility.

The company has already built more than 350 Figure 03 robots and increased production from one robot per day to one robot per hour in under 120 days. That is a 24x production improvement.

Figure says the factory now has dedicated lines for critical robot modules, more than 150 networked workstations, over 50 in-process inspection points, and more than 80 functional verification tests for each robot before approval.

The company also shared some early production numbers: more than 500 battery packs shipped, over 9,000 actuators produced, and battery-line first-pass yield at 99.3%.

The bigger story is data. More robots means more real-world testing, more failures discovered, more fleet learning, and more training data for Helix, Figure’s humanoid AI model.

Figure also revealed a new capability for Helix’s System 0: perception-conditioned whole-body control. In simple terms, the robot can now use camera perception to understand stairs, ramps, and uneven terrain, then move across them using a controller trained in simulation and deployed on real hardware without real-world fine-tuning.

Recursive Multi-Agent Systems

Recursive Multi-Agent Systems makes AI agents work as one loop instead of passing long text back and forth. “Latent space” means hidden number-form thinking. The system improved accuracy by 8.3%, ran up to 2.4x faster, and used far fewer tokens, or text pieces.

Co-Director: Agentic Generative Video Storytelling

Co-Director helps AI videos stay consistent from scene to scene. It uses agents like a film crew: one tests creative options, another checks errors. Its benchmark, GenAD-Bench, tests 400 ad scenarios, and Co-Director scored higher than other systems.

Stochastic KV Routing: Enabling Adaptive Depth-Wise Cache Sharing

Stochastic KV Routing cuts memory costs for language models. A KV cache is saved memory that helps AI avoid rereading past words. The method lets some layers share that memory instead of storing separate copies, often keeping quality while shrinking memory use.

Higgsfield Canvas: One Workspace for AI Videos: Higgsfield launched Canvas, a drag-and-connect workspace for AI images and videos. Instead of jumping between tools, creators can place prompts, images, video models, and edits in one shared board; “nodes” simply mean connected blocks in a visual workflow.

Qwen FlashQLA: Faster AI Under the Hood: Qwen released FlashQLA, a tool that makes some AI models run faster on Nvidia chips. A “kernel” is a tiny piece of code that tells the chip how to do work, and this one can speed up key model steps by about 2x to 3x.

DeepSeek Adds Vision, Raises the Price Pressure: DeepSeek is testing image recognition in its chatbot, meaning it may soon handle pictures, not just text. That matters because DeepSeek is already known for low prices, and adding vision could force rivals to cut costs or offer more for the same price.

Zuckerberg-Backed Biohub Bets $500M on AI Biology: Chan Zuckerberg Biohub is putting $500 million into AI models of the human body. The goal is to build better “simulations,” or computer versions of cells, so scientists can predict disease faster and test ideas before doing expensive lab work.

Anthropic May Be Valued Above $900B: Anthropic is weighing new funding offers that could value the company at over $900 billion. A “valuation” means what investors think a company is worth, and this would make the Claude maker one of the most valuable AI startups in the world.