1. OpenTelemetry Profiles Enters Public Alpha
The OpenTelemetry Profiling SIG has officially launched Profiles into public alpha. This establishes a unified industry standard for continuous production profiling, positioning it alongside existing standards for traces, metrics, and logs.
2. Google Releases Gemini 3.1 Flash Live Preview
Google has introduced the Gemini 3.1 Flash Live Preview, which now features configurable thinking levels. When set to high, the model scores 95.9% on the Big Bench Audio Speech to Speech benchmark, making it the second-highest scoring speech reasoning model.
3. Mistral Open-Sources Text-to-Speech Model
Mistral AI has released a new open-source text-to-speech model, providing the weights for free. The company claims the new speech generation model outperforms ElevenLabs in voice synthesis capabilities.
4. Cohere Launches Transcribe ASR Model
Cohere has released Cohere Transcribe, a state-of-the-art Automatic Speech Recognition (ASR) model designed for enterprise speech intelligence. The open-source voice model aims to simplify the extraction of actionable text from unstructured audio.
5. Intercom Debuts Purpose-Built Fin Apex 1.0 Model
Customer service platform Intercom has announced Fin Apex 1.0, a small, purpose-built AI model trained in-house. The company claims the model outperforms frontier models like GPT-5.4 and Claude Sonnet 4.6 specifically on customer service resolution metrics.
6. Cursor Details Real-Time RL for Composer
Cursor has shared insights into how it improves its Composer feature using real-time reinforcement learning. The team is extracting training signals from trillions of tokens of inference volume generated by users in the real world.
7. AI Rewrite of JSONata Saves $500K Annually
An engineering team successfully used AI to rewrite the JSONata API surface in a single day. The reimplementation reportedly saved the company $500,000 per year in operational costs.
8. Turbolite: SQLite VFS for Sub-Second S3 Queries
A developer has released Turbolite, an experimental SQLite Virtual File System (VFS) written in Rust. The tool is designed to serve cold JOIN queries directly from S3 object storage with sub-250ms performance.
9. Stripe Introduces Projects CLI
Stripe has launched Stripe Projects, a new command-line interface tool. The CLI allows developers to provision and manage Stripe services directly from their terminal.
10. Chroma Open-Sources Agentic Search Tools
Chroma has published its first open release on Hugging Face. The release includes a technical report and comprehensive details on training an agentic search agent.
11. VideoCUA Dataset Released for Computer-Use Agents
Researchers have open-sourced VideoCUA, a large-scale video dataset designed to advance human-level computer-use agents. Hosted on Hugging Face, the dataset contains 55 hours of footage, 6 million frames, and 10,000 human-demonstrated tasks across 87 desktop applications.
12. ATLAS Framework Pushes 14B Model Past Claude Sonnet on Coding
The Adaptive Test-time Learning and Autonomous Specialization (A.T.L.A.S) framework has achieved a 74.6% pass@1 rate on LiveCodeBench using a frozen 14B model on a single consumer GPU. The system uses constraint-driven generation and self-verified iterative refinement to outperform larger models like Claude Sonnet.
13. Tencent Open-Sources 7B Covo-Audio Model
Tencent AI Lab has released Covo-Audio, a 7-billion parameter end-to-end Large Audio Language Model. The architecture unifies speech processing and language intelligence, allowing it to directly process continuous audio inputs and generate audio outputs for real-time conversations.
14. Voxtral-4B-TTS Model Released on Hugging Face
A new text-to-speech model, Voxtral-4B-TTS, has been released with a demo available on Hugging Face. The 4-billion parameter model aims to compete directly with closed-source voice generation APIs.
15. Tech Companies Shift Toward In-House Open Models
Companies including Pinterest, Airbnb, Notion, Cursor, and Intercom are increasingly training and deploying open models in-house. Engineering leaders report that for many tasks, self-hosting and fine-tuning open models is proving better, cheaper, and faster than relying on proprietary APIs.
16. Swift 6.3 Released
Apple has released Swift 6.3, continuing the language's expansion across the software stack. The update maintains Swift's focus on strong safety guarantees and performance control for everything from embedded firmware to internet-scale services.