Models & Labs
TypeSafe launches Jev: non-generative AI for software
The Rundown AIhigh confidence
Why it matters
- →Jev offers a deterministic alternative to probabilistic LLMs for structured decision-making tasks.
- →Pricing at $42/B tokens is significantly lower than current frontier model rates.
- →Zero-hallucination design makes it suitable for high-stakes internal software logic.
TypeSafe has emerged from stealth with Jev, an AI system designed to make binary or categorical decisions within software rather than generate text. Co-founded by ChatGPT creator Diogo Almeida, the model operates at $42 per billion input tokens and delivers responses in 70-500 milliseconds, claiming to eliminate hallucinations by restricting outputs to preset options. The company positions Jev as a 'frontier-intelligence function call' for tasks like data sorting, record scoring, or jailbreak detection, aiming to replace expensive LLMs for routine judgment calls.
Read originalMore from The Rundown AI
More in Models & Labs
Models & Labsmodels
llama.cpp adds support for DFM Mimir 1B HRM architecture
llama.cpp now runs the DFM Mimir 1B model, a rare example of a hierarchical recurrent transformer that alternates between two distinct processing cycles. This isn't just another standard decoder; it requires handling aliased cache slots and looped graph builds, pushing the runtime's architectural flexibility further than usual. While the decode speed is roughly four times slower than a dense model of similar width due to these 128 block passes, local inference for this specific architecture was previously impossible on consumer hardware. This release effectively opens the door for experimenting with non-standard transformer topologies in open-source environments.
llama.cpp Releases
Models & Labscoding
llama.cpp b11007 enables CUDA graphs for MTP
This release quietly patches a major performance bottleneck in speculative decoding. By enabling CUDA graphs for Multi-Token Prediction (MTP), llama.cpp reduces kernel launch overhead on NVIDIA GPUs, making draft-token generation significantly faster. It also updates CUDA builds to support versions 12.8 and 13.3, keeping pace with the latest driver stacks. While KleidiAI is temporarily disabled on Apple Silicon, the focus here is squarely on squeezing more throughput out of existing hardware for local inference.
llama.cpp Releases
© TechCrunch AIModels & Labsother
Meta develops camera-free AI glasses 'Luna'
Meta is quietly addressing the privacy backlash against its Ray-Ban smart glasses by developing a new model called Luna that strips away cameras entirely. This device relies on six microphones and a side button to interact with Meta's AI chatbot and Muse agent, positioning audio-first interaction as a viable alternative to visual surveillance. By decoupling AI assistance from constant recording, Meta is testing whether users will adopt wearable AI if the social stigma of being watched is removed. It signals a strategic pivot toward privacy-preserving hardware to broaden adoption beyond early adopters comfortable with always-on cameras.
TechCrunch AI