Models & Labs
Typesafe AI launches Jev for high-speed classification
Sam Witteveenmedium confidence
Why it matters
- →Jev targets the latency bottleneck in agent workflows by prioritizing speed over complex reasoning.
- →It represents a niche strategy of using specialized models for classification rather than general-purpose LLMs.
- →Early demos suggest measurable performance gains for structured data extraction tasks.
Typesafe AI has introduced Jev, a new model architecture optimized for System 1-style rapid classification tasks. The company positions Jev as a solution for high-throughput scenarios where low latency is critical, contrasting it with traditional large language models that rely on slower, multi-step reasoning processes. Early demonstrations highlight its ability to handle choice, scoring, and entity extraction tasks with significant speed improvements over standard generative approaches. This release marks Typesafe AI's entry into the market for specialized inference models aimed at reducing computational overhead in automated pipelines.
Read originalThe story around this
TopicJev AI ModelCooling
Earlier coverage that leads up to this article, and what followed. Lines connect each piece to the closest one after it, converging here.
How we got here
- 1Scalable Enterprise AI Hinges on Agent Logic
Hugging Face Blog · June 1, 2026 · Background
- 2Together AI Hosts MiniMax M3 for Efficient Inference
Together AI Blog · June 2, 2026 · Background
- 3Benchmarking Open Models for Agentic Use
Hugging Face Blog · June 18, 2026 · Background
- 4Subquadratic claims breakthrough in LLM efficiency
MIT Technology Review AI · June 19, 2026 · Background
- 5OpenAI and Broadcom unveil LLM-optimized chip
OpenAI · June 24, 2026 · Background
- 6Scaling AI Agents with Redis Iris
Cole Medin · July 9, 2026 · Background
- 7ThunderAgent Boosts Agentic Inference Efficiency
Together AI Blog · July 29, 2026 · Background
- 8MIT Develops AI Method for Safety-Critical Tasks
MIT News AI · September 14, 2026 · Background
What happened next
- 9Typesafe AI Launches Jev Decision Model
Matt Wolfe · September 28, 2026 · Same story
- 10Ollama adds Jev-style decision models
Ollama Blog · September 29, 2026 · Related
- 11Jev as a Smart Router for Agent Harnesses
Sam Witteveen · September 29, 2026 · Related
- 12OpenAI launches Decisions API for fast agent classification
TechCrunch AI · September 30, 2026 · Same story
- 13TypeSafe AI raises $870M for non-text model Jev
TechCrunch AI · October 9, 2026 · Related
Follow this story
Open the full story →TypeSafe Launches Jev Judgment Model
7 developments
- Sep 17 · The AI Daily Brief
TypeSafe Launches Jev Judgment Model - Sep 18 · Sam Witteveen
Typesafe AI launches Jev for high-speed classification (This article)↳ Jev prioritizes System 1 reasoning for ultra-fast classification, targeting latency bottlenecks in agent workflows rather than chain-of-thou - Sep 18 · TechCrunch AI
TypeSafe AI launches Jev, a non-LLM model for automation - Sep 21 · Cole Medin
Jev: A Decision-Only AI Model for Agents↳ Jev abandons text generation to become a pure decision engine with sub-200ms latency and negligible costs. - Sep 21 · Fireship
Ex-OpenAI researcher launches Jev System 1 model↳ Ex-OpenAI researcher Diogo Almeida launches Jev System 1, claiming it is 200x faster and 400x cheaper than standard LLMs by stripping out te - Sep 22 · Duncan Rogoff
Jev AI Classifier: Speed vs. Hype↳ Jev excels at rigid schema-based routing but fails on open-ended generation or complex reasoning. - Sep 23 · Together AI Blog
Together AI: Fine-tune Jev-like classifier for $17↳ Together AI fine-tunes Qwen3.5 4B into a Jev-style classifier for $17 in 25 minutes.
More from Sam Witteveen
© Sam WitteveenModels & Labsagents
BAAI Releases AREX-2 Agent Model
BAAI’s AREX-2 shifts the burden of agent reliability from external orchestration to the model itself. Fine-tuned on Qwen3.8-27B, this open-weights model is explicitly trained to execute the try-check-retry loop internally rather than relying on a separate harness to manage failures. This architectural choice targets long-horizon tasks where external loops often break down under complexity. By embedding reflection and self-correction directly into the weights, it offers a more robust path for autonomous coding agents that need to recover from errors without human intervention.
Sam Witteveen
© Sam WitteveenOpen Sourceother
Ricursive AI Releases Ripple 65B for Chip Design
Ricursive AI has released Ripple 65B, an open-source model specifically trained to accelerate semiconductor design workflows. By focusing on a niche vertical rather than general-purpose reasoning, the company achieved competitive performance against major US open models while training for just $32 million. The architecture is described as thin and deep, optimized for the specific token distributions found in hardware description languages. This release signals that specialized, domain-locked models can outperform broad foundation models when the data distribution is sufficiently distinct and high-value.
Sam Witteveen
More in Models & Labs
Models & Labsother
llama.cpp b12177 adds CUDA 13 and ROCm 10
This release quietly closes the hardware gap for local inference by adding default support for CUDA 13.5 and ROCm 10.0 alongside existing CUDA 12 builds. AMD GPU users finally get first-class parity with NVIDIA, while Snapdragon mobile chips gain dedicated NPU acceleration paths. Apple Silicon KleidiAI is explicitly disabled here, a notable regression for Mac performance that suggests ongoing tuning issues. The inclusion of OpenVINO and SYCL backends further cements llama.cpp as the universal runtime for inference across nearly every major silicon architecture.
llama.cpp Releases
Models & Labsmodels
llama.cpp adds Kimi K3.1-Pro support
Kimi’s new K3.1-Pro model joins the llama.cpp ecosystem, bringing a distinct architectural twist to local inference. By replacing full-attention MLA layers with DSA (Dynamic Sparse Attention) and linear-attention KDA layers, this release tests how well sparse caching handles hybrid attention patterns. The implementation keeps indexer precision high to avoid quality loss during top-k selection, ensuring the model runs accurately on consumer hardware. This expands the library’s reach into newer Chinese open-weight models that rely on complex memory management.
llama.cpp Releases
Models & Labsother
llama.cpp b12182 release with CUDA 13 and ROCm 10
This release quietly cements llama.cpp as the universal inference runtime by adding default builds for CUDA 13.5 and ROCm 10.0 across Linux and Windows. The inclusion of ROCm 10 means AMD GPU users no longer need custom compilation to access modern acceleration, effectively closing the gap with NVIDIA's ecosystem support. While KleidiAI on Apple Silicon is disabled in this build, the broader platform coverage ensures that local inference remains accessible on virtually any hardware configuration without manual setup.
llama.cpp Releases