PRODUCTS

Mistral gives hands-free productivity a boost

A month ago, Mistral conquered speech-to-text with its Voxtral Transcribe 2 model family. Now, the AI lab is flipping the script with its first text-to-speech model.

On Thursday, Mistral launched its Voxtral TTS, an open-weight text-to-speech model that is three times smaller than the industry standard, more natural-sounding, and multilingual. The company claims that it's the best open-source text-to-speech model released to-date.

The model was built to be as natural-sounding as possible by understanding the context of what people say and modeling how people naturally speak, including pauses, rhythm, intonation, emotion and more. Some other highlights include: 

• Support for nine languages: English, French, German, Spanish, Dutch, Portuguese, Italian, Hindi, and Arabic. It can accept input in one language and output in another.

• Speed: Adapts to voices with references of just five seconds and offers low-latency, taking a voice sample of 10 seconds and 500 characters, and producing audio within 90 milliseconds

• On-device: 3 billion parameter model that fits in a smartwatch, a phone, or a laptop

• Human evaluations: In comparative human evaluation of Voxtral TTS and ElevenLabs v2.5 Flash, Mistral’s model achieved superior naturalness

While there are many other text-to-speech models on the market, Pierre Stock, VP of science and the first employee hired at Mistral AI, said the model’s overall efficiency sets it apart. 

“The model is optimal for both efficiency — including the cost, the speed, the ease of deployment – and state-of-the-art, conversational capability in which the model feels both super natural and conversational,” Stock told The Deep View. 

Some use cases, Mistral notes, include customer support, sales and marketing, real-time translation and personal voice agents. The model is available to test out in Mistral Studio playground, and a model with several reference voices is available as open weights on Hugging Face.

Since the dawn of computing, users have been dependent on a keyboard to get tasks done. As a result, one of the most exciting applications of generative AI is the ability to interact with applications more intuitively and hands-free via multimodal understanding, including voice. However, this poses many challenges, as seen by the failure of AI voice assistants such as Alexa and Siri to deliver really conversational and helpful experiences. To bridge the gap, more developments are needed in text-to-speech and speech-to-text, and Mistral is doing just that with this new model.

TOGETHER WITH AIRIA

One-Stop Enterprise AI At Your Fingertips

These days, it seems like there’s a different AI tool for every single business need – sales, marketing, creative, team lunch orders (okay, maybe not that one, but you get the idea). And while that’s all well and good, it can get a bit complicated… not to mention expensive. Which is all the more reason to opt for a platform like Airia that does everything, and does it well.

Airia’s enterprise AI platform gives you all the tools you need to up your productivity and innovation, maintain security and compliance, and manage it all in one simple, seamless hub. That means confusion, less crossed wires, and more “getting the job done” – and isn’t that what AI is supposed to be about in the first place?

Try a demo of Airia’s enterprise AI for yourself right here.

STARTUPS

Can SLMs solve AI's biggest roadblock in 2026?

When I talk to people putting AI to work in organizations right now, I typically hear two things: either inference costs are out of control or they are paying too much for tokens. Spoiler: those two things mean the same thing. The cost of running AI is too high, and AI agents are making it a lot worse. 

Here's the good news (or bad news, depending on your POV): If you're building AI agents, most of the tokens your agent consumes are going to a frontier model that's wildly overqualified for the job, and you're overpaying for it.

We're talking about tasks like:

• Classifying a support ticket 

• Extracting a date from an email 

• Rewriting a subject line

These are not tasks that require models with hundreds of billions of parameters. A small, fine-tuned model can handle them in milliseconds at a fraction of the cost.

In an exclusive interview with The Deep View, Rob May, CEO of Neurometric, shared how the company is betting big on this mismatch. 

On Wednesday, Neurometric launched its SLM Marketplace, a catalog of 115 task-specific small language models (SLMs). Each of these models is under 20 billion parameters and is fine-tuned to do one thing well. You can download any model for free, or let Neurometric host it with up to 100 million tokens per month at no charge. After that, it costs $2 per month per model.

"100 million tokens on one of these models costs us 40 cents," May told me. "So why would we charge for it?"

May says they are seeing a 25/75 split: roughly 25% of tasks in a typical enterprise AI workflow genuinely need frontier-level reasoning. The other 75% are structured, repetitive, and narrowly focused, such as classification, extraction, formatting, routing, scoring, and summarization. SLMs can handle those tasks.

The models in the SLM Marketplace span 14 categories, from accounting and finance to developer tools. Neurometric expects to scale to 5,000 models by year's end. For tasks not yet covered, an Auto-SLM Creator can generate custom models from a plain-language description.

One insight that stood out from my conversation with May was that the biggest challenge with small models is keeping them on track. Larger models handle this naturally, but SLMs need what May described as "micro harnesses," or checkpoint prompts that remind the model what it's doing and what comes next. It's part of an emerging discipline called harness engineering.

One of the common themes I've been hearing over and over again across the AI ecosystem in 2026 is that we're moving past the idea of one model to rule them all. There's a larger diversity of models than ever, and they have different strengths and weaknesses. And it's not just LLMs that are ruling the day. Small models, domain-specific models, and now task-specific models are finding a sweet spot because of their cost and performance advantages. We're moving into a multi-model world where the smartest architecture is the right model for the right task at the right cost.

TOGETHER WITH TELEPORT

AI Agents Need Infrastructure Guardrails

AI agents are starting to move beyond copilots and into real execution across infrastructure. That shift introduces a new layer of risk, as most identity systems still rely on static credentials and human assumptions that don’t translate to autonomous actors.

Teleport’s Agentic Identity Framework brings identity-based access, short-lived credentials, and enforceable policy controls to infrastructure access across Kubernetes, databases, and services. It’s designed to help teams run AI systems securely in production without exposing secrets or over-permissioning access.

Explore the platform.

CULTURE

Grok chases AI video rivals in post-Sora market

As OpenAI winds down its video effort, xAI wants to fill the void. 

Following OpenAI’s Tuesday announcement that it would scrap Sora, its generative video app, to focus on its enterprise AI work, Elon Musk-owned xAI announced plans to step up Grok Imagine, the lab’s video generation tool. 

In a post on X, Musk said that the company is “doubling down” on AI video, claiming that “the next Grok Imagine release will be epic.” 

Though OpenAI’s decision to wind down Sora and end its $1 billion, three-year partnership with Disney for licensing was a surprise development, it’s indicative of OpenAI’s efforts to cut down on side projects and refocus on core enterprise products that drive revenue, especially as it eyes an IPO and likely feels pressure from the success of rival Anthropic.

And while xAI may aim to fill the gap left by OpenAI's withdrawal of the Sora app, xAI's image generation tool remains shrouded in controversy. 

• The company has faced a litany of lawsuits for its image generator creating photos portraying people in sexualized scenarios without their consent, including from a group of teenagers in Tennessee who proposed a class-action lawsuit claiming that Grok generated images of them in debasing scenarios. 

• Additionally, the company has faced EU regulatory probes regarding its image generator, and is facing a lawsuit from the city of Baltimore for the image generator’s safety protocols.  

A recent study from advocacy group Common Sense Media deems Grok to be a risk for users under 18, citing “frequent generation of sexual, violent, and inappropriate material.”

Though Sora was still in its early days, its loss leaves a gap in the AI video market. And while Musk may see this as an opportunity, xAI’s tech may not be up to the task of taking Sora’s place. For one, there are plenty of other competitors, ranging from startups like Runway or Synthesia to tech giants like Google's Veo and ByteDance’s Seedance. With a wealth of options, entertainment companies like Disney may be more inclined to work with companies that have cleaner reputations than xAI. 

LINKS

• Google launches TurboQuant for LLM efficiency at high accuracy

• Disney ends three-year, $1 billion OpenAI licensing deal 

• Granola raises $125M in Series C funding at $1.5 billion evaluation

• Meta uses AI to update shopping experience across apps

• Lucid Bots, a window-washing drone company, raises $20 million

• Google rolls out Gemini for Google TV updates

• Legal AI startup Harvey raises $200 million at $11 billion valuation

• Google Ads: Veo is now available in the advertising platform

• Sora app: The app is being discontinued by OpenAI, with timeline coming

• Freepik: Relight launched Freepik, giving users more lighting control

• Grok: Grok Imagine multi-image to video and video extension available on API

• Anthropic: Security Labs Engineer

• Qualcomm: Senior Data Lead – Autonomous Driving Stack 

• Zoox: Machine Learning Engineer - Multi-Modality Foundation Model

• Tesla: Kernel Optimization Software Engineer, AI Hardware

GAMES

POLL RESULTS

Have you ridden in a self-driving cab such as Waymo, Zoox, or Robotaxi?

Yes (13%)
Not yet, but I'd like to (47%)
No, and I don’t intend to (37%)
Other (3%)