Google Cloud is now hosting OpenAI models

Recursive just released Flint, an Apache-2.0 licensed coding Mixture-of-Experts model. It packs 40 billion total parameters and a massive 262K-token context window, explicitly designed to run locally as a complement to heavier cloud-based agents.

Relying entirely on frontier cloud APIs for your dev workflow is a recipe for latency and downtime. Flint gives you a highly capable, local alternative that can swallow entire codebases in a single prompt without sending your proprietary logic over the wire.

If the performance holds up to the specs, this is a clear win for local-first development. The cloud providers lose their monopoly on your keystrokes, and developers get a reliable, private fallback that actually understands context.

Modal, the AI infrastructure platform, just crossed the $1 billion annualized run rate milestone. On the back of that growth, they secured a $750 million Series D funding round, pushing the company's valuation to a staggering $16 billion.

While foundation model builders burn billions chasing theoretical intelligence, the infrastructure providers are quietly minting cash. Modal's massive Series D highlights an insatiable, entirely practical demand for scalable AI compute that just works.

The winners here are the picks-and-shovels providers. If you are building AI applications, Modal's war chest means more stable, specialized infrastructure is coming your way. The losers are the generalist cloud providers who thought they could own the entire AI deployment pipeline by default.

Mirendil claims to have built its first base model using a single researcher over just 3.5 months. More surprisingly, the company states this entire pretraining run cost under $10 million in compute, demonstrating a severe drop in the resources required for large model development.

The current consensus dictates that training competitive base models requires massive teams and nine-figure compute budgets. Mirendil's claim directly challenges that assumption, suggesting that algorithmic efficiency and focused execution can substitute for brute-force capital.

We need to see independent verification of the model's actual capabilities before rewriting the rules of AI economics. But if validated, this drastically lowers the barrier to entry for training custom models from scratch.

Perplexity and NVIDIA introduced JOLT, a joint teacher-student post-training method. The technique addresses the known weaknesses of self-distillation by training the teacher model using outcome rewards, while actively penalizing it for excessive divergence from the student model's behavior.

Distilling massive models into deployable sizes often results in a loss of reasoning quality. By forcing the teacher and student to learn in tandem, JOLT improves performance on benchmarks and creates a more stable pipeline for creating efficient, production-ready models.

This is a highly practical development for teams trying to serve high-quality outputs without paying frontier-model inference costs. Better distillation means smaller, cheaper models that actually perform their intended tasks.

DeepSeek released DeepInfer, a new framework that exposes advanced production scheduling techniques for model inference. The system combines speculative decoding, dynamic prefill sizing, and other optimization methods to maximize throughput.

Serving large language models at scale is an infrastructure nightmare. By open-sourcing the scheduling logic they use internally, DeepSeek is giving independent developers the tools to drastically reduce latency and compute costs in production environments.

This is a massive win for infrastructure teams. You no longer have to reverse-engineer how the major labs achieve their inference speeds. Implementing DeepInfer's dynamic prefill sizing should be the top priority for anyone running models at scale.

Researchers discovered that when large language models are asked open-ended questions about culture, they exhibit a distinct bias toward countries like Japan. This cultural favoritism does not appear in base models; it emerges specifically after supervised fine-tuning.

We treat fine-tuning as a mechanism for aligning models with human values, but we rarely ask whose values are driving the steering wheel. This research proves that the alignment process inadvertently injects specific cultural preferences into the model's worldview.

If you rely on LLMs for global content generation or unbiased cultural analysis, you need to account for this lean. The models are not objective observers; they are reflecting the specific, localized preferences of the annotators who trained them.