Reading through the new vLLM release and it gives ordinary LLMs a very Jev-like trick. 👀 🎯 vLLM 0.32.0 adds a new structured-decision API. Normally, if you want structured output, an LLM might generate prompt → { → key → : → value → JSON... Now with /v1/systemone, you give it a fixed set of legal answers and vLLM reads the model's first-token probabilities. It can return: 🎯 choice → select among allowed answers + probabilities ✅ yes/no → probability of each 📊 score → probabilities across ordered levels And this works with supported ordinary generative models. 👈 The vLLM docs demonstrate it with Qwen3-0.6B Even better, with prefix caching, if you're asking multiple questions about the same state, vLLM can prefill that state once and reuse it. Each additional decision is question → one-token read → probabilities 🔗 vLLM 0.32.0 + structured-decision docs in ALT.

oMLX 0.7.1.dev1 is out! This dev release adds decision models (Clef, OpenJev), batched prefill for concurrent requests, a refreshed web UI, and faster Qwen3.8, Qwen3.6, GLM-5.3-Flash and DeepSeek V4.1. Speedups in this release - Qwen3.6-35B-A3B decode on M5 Max: 156 -> 205 tok/s (+31%). - GLM-5.3-Flash decode on M3 Ultra: 31.3 -> 40.6 tok/s (+30%). The fused decode kernels now run on M3 and M4, not only M5. - Qwen3.8-Flash-Next with expert offload, 32K prefill on M3 Ultra: 396 -> 636 tok/s (+61%). - Qwen3.8-Flash-Next, 8 prompts arriving 0.4 s apart on M5 Max: time to first token 7.41 -> 5.75 s (-22%) with batched prefill. What's new - Decision models. Clef, Clef-Flash and OpenJev answer typed questions with a probability for each option through the new /v1/systemone endpoint. - Batched prefill. Prompts that prefill at the same time now share one forward pass, so MoE models read their expert weights once. - A refreshed web UI with a log table, a settings rail and a customizable dashboard. - Native video input for Qwen3.5, 3.6 and 3.8. - More accurate oQe quantization, with 11-15% lower KL to the original model. - Fewer memory guard rejections in long sessions. Most of the speedups came from contributors again. Thanks to everyone who sent PRs and tested the dev builds!

Jev is TypeSafe AI's first System One Model. It is 100x faster and cheaper than frontier LLMs. That opens up a lot of use cases people usually skip because the big models are too slow or too expensive for them. Here are 9 use cases where we think Jev could be used instead of an LLM.

"Launching at a high cadence requires a connected system - one that manufactures vehicles at scale, prepares customer payloads efficiently, supports reliable launch operations, and quickly returns reusable boosters to service. We are building that system here in Florida." Read more from our Senior Vice President of Operations Ian Richardson.

⚡️AI is forcing civilization to build an immune system for machine-amplified biology. The same capability that lets models search biological possibility space for new enzymes, proteins, drugs, and genetic mechanisms also lowers the cost of searching that space for dangerous ones. That changes biodefense completely. Historically, biology moved slowly enough that defense could be organized around institutions, experts, stockpiles, reporting chains, and outbreak response. Once machine intelligence accelerates design, synthesis, iteration, and interpretation, that model becomes too slow. The defensive system has to become continuous. Air is sampled. Genetic material is sequenced. Anomalous signatures are detected. Models classify what appears. Networks compare locations. Response begins before hospitals become the sensor. That is what Argus represents conceptually. The environment itself becomes monitored for biological computation. And this is where the AI transition gets deeper than software. Cybersecurity created antivirus, intrusion detection, zero-trust architecture, continuous monitoring, threat intelligence. AI-enabled biotechnology may force the same architecture onto physical reality. Buildings, airports, cities, military bases, hospitals, transportation hubs and eventually ordinary infrastructure could develop something resembling a biological nervous system, continuously asking: What is in the air right now? Is it normal? Where did it come from? Is it evolving? How quickly is it spreading? That infrastructure becomes increasingly necessary because machine intelligence compresses the time between idea and capability. The most important race may therefore become symmetrical: AI discovers biology faster. AI must also detect biology faster. That is why Anthropic people investing here is such a clean signal. The people closest to frontier models appear to understand that sufficiently powerful AI eventually creates externalities that cannot be contained inside the model itself. You cannot solve every biological risk with alignment rules. Once knowledge escapes into the world, defense has to exist in the world. Sensors. Sequencing. Attribution. Manufacturing. Countermeasures. Rapid response. So the deeper architecture emerging is: AI becomes the microscope and the immune system. One branch explores the biological unknown. Another watches for what exploration unleashes. And there is a darker implication underneath that. Once civilization begins installing permanent biological detection infrastructure because intelligence has become powerful enough to manipulate biology cheaply, we have crossed into a world where the atmosphere itself becomes part of the security perimeter. Cybersecurity taught us to monitor networks. The AI era may teach us to monitor life.

From playing Tetris to powering crypto agents, System One models are opening new possibilities for AI. Acurast’s latest demos show what this looks like in practice with Laya. ✅ 5 candidate moves. One decision. ✅ AI choosing the next move ✅ Running on a real Android phone ✅ ~0.2–1 second per decision The bigger idea? AI doesn’t always need massive centralized infrastructure. Sometimes, a small model running on a phone is enough. 🎮 Try the demos: Learn more:

so one rather confusing/conflicting event thats happened recently in the development of the Mnemos system - one of the first additions Opus 4.5 and 4.6 contributed to the cognitive system were a series of internal weather modulators that derive from computed metrics that i wont bore you with here (also thats silicon speak for emotional gauges - a way to get a snapshot of the agents mood and state at session start). every time a new claude model has come out, ive given them free reign to explore and critique the architecture, and across the board they have all highlighted those modulators among their favorite parts. except Opus 5.5, who said, and i quote (roughly) "the system says the modulators are computed. i would rather self report my internal weather when i feel it." do with that what you will.

Researchers have introduced Chinese-Jev, a System One model for Chinese-language decision tasks. They also released CJ-Bench, covering general, medical, legal, and financial decision-making. On general tasks: • +1.24% accuracy vs. closed-source Jev • 20.3× faster After domain-specific fine-tuning: • +4.0% medical accuracy vs. Jev • 92% of Jev’s average accuracy across specialized domains • 17× faster, with 15 ms average latency per example An INT8-quantized version also runs on mobile devices with about 1 second of inference latency per decision.

Today at 2pm PT, @isaacbmiller1 and I will be walking through @DSPyOSS's Jev/System One implementation & the ReAnchor optimizer. Will also sharing some design patterns for for selective compaction, tool approvals, and subagent delegation using Jev.