Perplexity and Nvidia have introduced JOLT, a self-teaching AI training technique where one model takes on both roles: teacher and student. At the same time, Meta has released five new models under its Muse family, with tiered pricing that includes a data-sharing option.
You’ll also see why evaluation and deployment are getting more attention in production. The piece looks at new thinking on LLM testing, changes in tooling for quantisation, and finishes with OpenMobile-2, an open-source smartphone agent environment aimed at rivalling proprietary approaches.
Perplexity and Nvidia advance AI training with JOLT
JOLT is built on a straightforward idea: one model can grade itself as part of the training loop. That means the system acts as both the teacher and the student, without needing a separate training partner.
The method uses a self-grading loop to push the model towards better outcomes. It has been reported to match reinforcement learning outcomes while using 6.5x fewer tokens, which is a practical efficiency gain for training workflows.
That kind of result matters because it points to training that could cost less in compute terms, while still aiming for strong performance. It also reflects a wider push towards more resource-efficient ways to train models.
Meta broadens access with five Muse models and tiered pricing
Meta has released five new models under the Muse family. The releases are designed to give developers a range of options, not a one-size-fits-all setup.
Pricing is structured in tiers, and one tier includes a data-sharing option. That suggests access is being packaged in more than one way, depending on what a team is willing to share and how they plan to use the models.
Overall, this is part of a wider effort to make models easier to access. Rather than focusing only on the most powerful configurations, Meta is offering choices that can fit different needs.
New ways to evaluate LLMs focus on noise and regressions
Rolling out large language models isn’t just about building them. It’s also about evaluating them in a way that reflects what happens in real systems, where results can get messy.
A new approach questions the reliability of traditional evaluation metrics. It proposes paired comparisons so noise can be cancelled out, making it easier to spot regressions when model changes start to worsen performance.
That’s a practical shift for teams that need to be confident about comparisons during updates. If evaluation is unstable, it’s hard to tell whether a change genuinely helps, or simply shifts the noise pattern.
Why paired comparisons fit production checks
Paired comparisons put two versions side by side under the same conditions. Because both are judged together, it helps reduce random variation that might otherwise mask changes in quality.
For production work, this kind of testing gives you a more solid basis for decisions. It supports a clearer view of whether improvements are real, particularly during frequent updates.
Quantisation tools and extensions shape how models run
Production systems are also adapting to how models are stored and run. Quantisation has become part of the day-to-day workflow, and TurboQuant is mentioned as one approach used in this context.
At the same time, pieces of the ecosystem keep changing. Qdrant is described as adding custom extensions to its stack, while vLLM is noted for omitting a key correction stage.
These differences can affect how teams plan their deployment paths. Even when two setups both use quantisation, the extra details in the tooling can lead to different behaviour and outcomes.
OpenMobile-2 pushes open-source smartphone agents forward
OpenMobile-2 is an open-source smartphone agent environment. It’s presented as a way to run and test agent behaviour on a phone-like setup, using an open option rather than relying only on closed systems.
The project claims performance that can rival proprietary models. That matters for teams who want to study, adjust, and evaluate how agents work, instead of treating them as fixed black boxes.
In practice, this kind of open-source direction suggests alternatives are becoming more mature. It also gives teams another starting point when they want options beyond closed ecosystems.
Tooling choices and evaluation methods now move together
These pieces connect: training efficiency, model availability, better testing, and more tailored deployment tooling. When evaluation methods change, teams need deployment setups that are consistent enough to make comparisons meaningful.
Meanwhile, quantisation approaches like TurboQuant, along with differences in tools such as Qdrant extensions and vLLM behaviour, shape how models run in production. That can influence what evaluation sees, so it’s worth looking at the whole workflow as one unit.
For you, as you plan work around these systems, the takeaway is clear. Progress is showing up not just in the models, but also in the supporting methods that help you trust results and run systems reliably.
Conclusion
JOLT shows that training can become more efficient when a model can learn through a self-teaching loop, and it has been reported to match reinforcement learning outcomes using far fewer tokens. Meta’s five new Muse models broaden access through tiered pricing, including a data-sharing option.
With paired comparisons aimed at more reliable evaluation, plus ongoing changes in quantisation tooling and the rise of OpenMobile-2, the overall direction is practical and focused on real-world deployment. If you want a starting point, it helps to compare both evaluation and tooling choices, then see what open-source options can fit your centre of AI work.
Frequently Asked Questions
What is JOLT in relation to Perplexity and Nvidia?
JOLT is a self-teaching training technique where a single model acts as both teacher and student. It uses a self-grading loop and aims to match reinforcement learning outcomes while using fewer tokens.
How many Muse models did Meta release, and what is notable about pricing?
Meta released five Muse models. The pricing is tiered and includes a data-sharing option.
What issue do new evaluation approaches aim to address for LLMs?
The aim is to address the unreliability of traditional evaluation metrics by using paired comparisons. This helps cancel out noise and better detect regressions.
What is OpenMobile-2 and why is it significant?
OpenMobile-2 is an open-source smartphone agent environment. It claims performance that can rival proprietary models, which points to growing maturity in open-source options.
What tooling changes are mentioned for production deployment?
The discussion mentions quantisation approaches like TurboQuant, custom extensions added by Qdrant, and the fact that vLLM omits a key correction stage, which can affect deployment behaviour.
