AIModel.fyi

Muse Glimmer 30B is a new 30-billion-parameter causal language model from the Meta Superintelligence Lab, distilled directly from Muse Spark.Released under the permissive Apache 2.0 license, it is purpose-built to execute complex, autonomous agentic workflows entirely on consumer hardware without relying on cloud infrastructure.

MiniMax H3 is a powerful new 33B parameter omni-modal generative system designed for unified understanding and generation of text, images, video, and audio. It is capable of generating highly synchronized videos with native 32kHz stereo audio, supporting up to 15 seconds of runtime and up to 2K resolution.

Role-stratified per-field conformal risk control calibrates LLM tool calls by semantic argument role rather than certifying each action as one aggregate object. In AgentDojo and InjecAgent, the method assigns separate thresholds and budgets to target, credential, command, selector, control, and content fields, matching certification to where an injection can cause harm.

Laguna S 2.1 is a new 118B parameter Mixture-of-Experts (MoE) model purpose-built for agentic software engineering and long-horizon tasks. Sitting comfortably between the XS and M.1 models in the Laguna family, it achieves high efficiency by activating just 8B parameters per token. More info is available on AImodels.fyi here! Massive 1M context window: Can process up to 1,048,576 tokens, making it ideal for ingesting large codebases and managing extended debugging sessions.

Large language models are trained to be generalists. They solve physics problems, write poetry, recall historical facts, and debug code using the same neural machinery. This means a model must dedicate precious computational resources to tasks that don’t really need reasoning at all. When asked “What is the capital of France?” the entire transformer backbone engages just to retrieve a fact.

Modern language models can refine their reasoning by looping back through their own computation, repeatedly applying the same layers to polish an initial answer. In theory, this is powerful. A model that thinks twice should produce better code than one that thinks once. But practice has a stubborn cost: if you loop sequentially, each additional pass multiplies both latency (time to first token) and memory usage (the KV-cache that stores what the model has attended to).

I bet you can imagine a researcher who runs an experiment, fails, and then forgets everything that led to the failure. The next attempt starts fresh, with no memory of what was tried or why it didn’t work. Over hours or days of research, this researcher would waste enormous effort revisiting dead ends, trying contradictory approaches, and never building cumulative knowledge about the problem. This is roughly how current AI agents approach autonomous research tasks.

Coding agents have gotten remarkably good at fixing bugs. The benchmark suites designed to measure this capability, like SWE-bench, keep pushing higher success rates. But something crucial is being obscured by these overall improvement metrics: we have no idea which specific skills are actually driving the gains. When an agent successfully resolves a bug, that success comes from at least three distinct capabilities working together.

Scientists spend enormous time hand-crafting publication-quality figures, yet every automated system in existence handles only one figure type at a time, producing static images that cannot be tweaked. The assumption underlying this limitation is straightforward: throw more data and model capacity at the problem, and eventually one system will master them all. This assumption is wrong, not because we lack compute or data, but because it misunderstands what a figure actually is.

SkillOpt: Executive Strategy for Self-Evolving Agent Skills Agent skills—the instructions and guidelines that govern how AI models behave when solving problems—exist in an awkward middle ground. They’re either hand-crafted once and frozen, generated fresh each time without learning, or loosely self-revised without any real feedback mechanism. None of these approaches behaves like actual optimization. Compare this to how we train neural networks.