1. Google Moves Gemini 3.5 Transcribe to General Availability and Launches Gemini 3.8 Live
Google has expanded its voice and audio capabilities by moving Gemini 3.5 Transcribe out of public preview into general availability and introducing the new Gemini 3.8 Live model. These updates provide developers with more stable, production-ready tools for real-time voice applications, following the initial preview of the Transcribe model announced in August and the broader Gemini Live API launch in July.
- • Gemini 3.5 Transcribe is now generally available, following its public preview launch on August 26, 2026.
- • Google introduced Gemini 3.8 Live, a new model optimized for real-time voice interaction.
- • The release provides developers with production-ready tools for low-latency voice recognition and live conversational agents.
Developers can now deploy production-grade real-time voice applications using the finalized Gemini 3.5 Transcribe and the newly available Gemini 3.8 Live model.
2. Prior Labs Releases TabPFN-3.5 Tabular Foundation Model
Prior Labs has launched TabPFN-3.5, a major upgrade to its tabular foundation model that scales the parameter count to 220 million. Designed to run on raw data with default settings, the model performs classification and regression tasks instantly without requiring per-dataset training. The release includes a faster 84M parameter version, an enterprise-only Plus version with native text handling, and a Thinking version that utilizes extra inference compute.
- • Prior Labs has released TabPFN-3.5, a 220-million-parameter tabular foundation model that performs predictions in a single forward pass.
- • The model achieved a score of 0.375 on the Otto Group Product Classification Challenge, beating the original winning Kaggle solution.
- • TabPFN-3.5 holds first place across seven tabular benchmarks, including TabArena and ScoringBench.
- • The model supports up to 1 million rows and up to 20,000 features (recommending 6,000).
- • Open weights are available for research, evaluation, and Kaggle use, while production deployment requires a commercial license or the Prior Labs API.
Developers can perform highly accurate tabular classification and regression in a single forward pass without needing to train or tune models per dataset.
3. Nums AI Releases Causilo Tabular Foundation Model
Nums AI has introduced Causilo, a tabular foundation model pretrained entirely on synthetic data. Causilo integrates directly into existing Python workflows via a standard scikit-learn interface, with its code licensed under Apache-2.0. By employing cross-attention, the model maintains linear computational complexity relative to the number of features, making it highly efficient for large-scale tabular datasets.
- • Nums AI has released Causilo, a pretrained tabular foundation model for classification and regression tasks.
- • The model features a scikit-learn interface, with Apache-2.0 code and pretrained weights available on Hugging Face.
- • Causilo holds the highest Elo among single models on the TabArena benchmark for both classification and regression.
- • The architecture uses a three-phase process (refinement, compression, and in-context learning) and cross-attention to keep computational costs linear to feature count.
- • The model requires Python 3.10 to 3.12 and PyTorch 2.13 or newer, and is free for research and evaluation.
Developers can integrate a high-performing, open-source tabular model into their Python pipelines using a familiar scikit-learn interface.
4. Knowledgator Releases GLiFormer Schema-Conditioned Encoder
Knowledgator Engineering has released GLiFormer, an open-source encoder framework designed to handle multiple extraction tasks without generating text tokens. By scoring compatibility against provided schemas and anchors directly, the model avoids the latency of traditional generative LLMs. The Apache 2.0-licensed checkpoints are available in Base (264M) and Large (575M) sizes, offering a fast and cost-effective alternative for structured data extraction.
- • Knowledgator Engineering has released GLiFormer, a schema-conditioned encoder framework for information extraction.
- • The model performs named-entity recognition, text classification, relation extraction, nested JSON structuring, and text embeddings using a single encoder.
- • GLiFormer Large (575.6M parameters) achieved an F1 score of 91.10 on nested JSON extraction benchmarks.
- • The model checkpoints are released under the Apache 2.0 license and are compatible with both CPU and GPU hardware.
- • GLiFormer-base (264.2M parameters) reported a median latency of 69 ms on an NVIDIA RTX PRO 6000 Blackwell GPU.
Developers can perform high-accuracy schema-conditioned information extraction and nested JSON structuring without the latency and cost of generative LLM tokens.
5. Ant Group Releases Ling-3.0-flash-Fin Open Weights Model
Ant Group has released Ling-3.0-flash-Fin, a specialized financial model built on the Ling-3.0-flash architecture. The model is designed to assist with domain-specific tasks such as source checking, report writing, and spreadsheet creation. While it demonstrates high business knowledge accuracy, evaluations show it also carries a higher hallucination rate compared to the vision-language variant of the same architecture.
- • Ant Group has released Ling-3.0-flash-Fin, an open-weights, finance-focused model under the MIT license.
- • The model features a Mixture-of-Experts architecture with 124B total parameters and 5.1B active parameters per token.
- • It supports a 256K token context window and is limited to text-only input and output.
- • The model is available on OpenRouter, including a rate-limited free endpoint.
- • Ling-3.0-flash-Fin scores 23 on the Artificial Analysis Intelligence Index, matching MiniMax-M2.7 while using half the active parameters.
Developers building financial applications can leverage a specialized, MIT-licensed open-weights model with a 256K context window for report writing and valuation tasks.
6. TypeSafe AI Launches Jev System One Model for Fast Decisions
TypeSafe AI has launched early access to Jev, a specialized "System One" model built for direct integration into software workflows. Unlike traditional LLMs that generate free-form text, Jev is designed exclusively for structured outputs and choice selection. The architecture is optimized to eliminate hallucinations entirely while operating at speeds up to two orders of magnitude faster than conventional models.
- • TypeSafe AI has introduced Jev, a new class of System One Model designed for fast, structured decisions in software.
- • The model is optimized for structured outputs and is stated to be incapable of hallucination.
- • Jev is claimed to be two orders of magnitude faster and more efficient than existing large language models while maintaining comparable intelligence.
- • The model is currently available in early access.
Developers can access a new class of highly efficient, hallucination-free models optimized specifically for fast, structured decision-making in software.
7. Google Launches Google Home Model Context Protocol Server
Google is opening up its smart home ecosystem to developer-built AI agents by launching an early access Model Context Protocol (MCP) server. This integration allows external agents, including Claude and ChatGPT, to query device states, review camera feeds, and control smart home hardware using natural language. To get started, developers must configure a Google Cloud project to interface with the Home MCP server.
- • Google has launched early access to a Model Context Protocol (MCP) server for Google Home.
- • The integration allows third-party AI agents like Claude and ChatGPT to control connected smart home devices and access camera summaries.
- • Setup requires developers to create and configure a Google Cloud project for Home MCP.
- • The feature is initially available to Google Home Premium Advanced users in the US for $20 per month or $200 per year.
- • Google has implemented safety protections, including rate limits and a restriction preventing agents from unlocking doors.
Developers can now build AI agents that directly interact with, monitor, and control Google Home smart devices using the Model Context Protocol.
8. Stanford Researchers Release Paper2Agent Pipeline
Stanford researchers have introduced Paper2Agent, an open-source pipeline that automates the process of turning scientific papers and their code repositories into Model Context Protocol (MCP) servers. The system manages environments, extracts tools, and runs a strict validation gate to ensure the generated agent's outputs match the original paper's figures and numerical results. The MIT-licensed tool allows developers to quickly build highly specialized scientific agents.
- • Stanford researchers have released Paper2Agent, a system that converts research papers and codebases into executable MCP servers.
- • The pipeline features a strict validation gate requiring numerical results to match within 3% and figures to match by perceptual hash.
- • In scale tests on 100 bioRxiv papers, 74 papers were successfully converted, with 593 of 599 tools passing validation.
- • The software is MIT-licensed, and prebuilt servers are hosted on Hugging Face Spaces.
- • Creating the AlphaGenome agent took 45 minutes and cost $14, with all 22 tools passing validation.
Developers can automatically convert scientific papers and their associated codebases into fully validated Model Context Protocol servers for AI agents to execute.
9. Ory Agent Security Launches Harness-Layer Governance for AI Agents
Ory Agent Security has introduced a vendor-agnostic security solution designed specifically for governing AI coding agents. By operating at the harness layer, the tool intercepts and evaluates actions before they reach external gateways. The system is compatible with 11 major AI coding agent harnesses, providing developers with a concrete way to implement identity and access management for autonomous agents.
- • Ory Agent Security has been released to govern AI agent execution at the harness layer.
- • The tool is vendor-agnostic and operates before any gateway calls are made.
- • Ory Agent Security is compatible with 11 major AI coding agent harnesses.
- • The solution is designed to address security challenges as AI agents increasingly outnumber human users.
Developers can secure and govern their AI coding agents by executing security policies at the harness layer before making external gateway calls.
10. Cloudflare Deploys 'Disallow AI Training' Setting Following September Deadline
Cloudflare has implemented the 'Disallow AI Training' setting, fulfilling the September 15 deadline for crawler separation established in July. This new feature allows site owners to block mixed-use crawlers from training AI models while permitting them to continue indexing content for search. The rollout replaces previous bot management tools with a unified 'Bot Preference Sync' system and includes commitments from major operators like Google, Apple, and Microsoft to honor the new opt-out setting.
- • Cloudflare has officially launched the 'Disallow AI Training' setting, meeting the September 15 deadline for crawler separation.
- • The feature enables site owners to block AI training by mixed-use crawlers while maintaining search discoverability.
- • Major operators, including Google, Apple, and Microsoft, have committed to honoring the new opt-out setting.
- • The update replaces legacy bot management tools with a new 'Bot Preference Sync' system.
- • Cloudflare plans to introduce further granular controls for AI-generated summaries by early next year.
This release provides the final mechanism for site owners to enforce the crawler separation policy, ensuring content is protected from unauthorized AI training without sacrificing search engine rankings.
11. Openjev Project Offers Open-Source Single-Pass Choice Selection
The openjev project offers an open-source, MIT-licensed alternative to commercial choice-selection models like TypeSafe's Jev. By utilizing an option-attention head to score inputs in a single pass, the model avoids the latency of standard token generation. The repository supports training with a default byte encoder or a frozen pretrained encoder from Hugging Face, such as Qwen2.5-0.5B, and is reported to be about 100 times faster than a small decoder forced to write 400 tokens when evaluating eight options.
- • The openjev project provides an independent, MIT-licensed starter model designed to choose among a changing list of text options in a single pass.
- • The model is reported to be approximately 100 times faster than a small decoder forced to write 400 tokens when evaluating eight options.
- • It uses an option-attention head to score inputs, converting options into query vectors that assign attention weights to context tokens.
- • In local experiments, the one-pass scorer reached approximately 98% accuracy on synthetic menus and 29% accuracy on Wikispeedia next-click data.
Developers can implement ultra-fast, single-pass classification and option selection in their applications without relying on slow token-generation decoders.
12. GSQ-RCO Quantization Further Optimizes Qwen3.8-Flash-Next for Local GPUs
Building on previous optimization releases for the 125B Qwen3.8-Flash-Next model—including NVIDIA's NVFP4 and AtomicChat's SSD offloading—the new GSQ-RCO quantization method provides an additional path for local deployment. By optimizing Q2_0 and IQ3_XXS variants, this method allows the model to run on 12GB VRAM hardware while maintaining near-baseline performance, offering an alternative to existing quantization and offloading strategies.
- • GSQ-RCO quantization reduces Qwen3.8-Flash-Next model size to 68-76GB.
- • The Q2_0 variant delivers 6.2x better prompt throughput and 1.9x lower latency than IQ2_XS.
- • The IQ3_XXS variant achieves 15 tokens/second output on a 12GB RTX 5070.
- • Performance matches the base model on AIME25 and remains within 0.51 points on GPQA-Diamond.
Developers now have an additional, highly efficient quantization option for running the massive Qwen3.8-Flash-Next model on consumer hardware, complementing previously released NVFP4 and SSD-offloading techniques.
13. vLLM Implementation Offloads Qwen3.8-Flash-Next KV Cache to RAM
Developers can now run Qwen3.8-Flash-Next with a 1 million token context length on vLLM by offloading the KV cache to system RAM. This approach leverages the model's unique architecture, which combines 12 KV cache layers with 36 gated delta-net layers. By utilizing the QSA mechanism to index and read a maximum of 2,048 positions per decode step, the system maintains high throughput even when the model's quantization barely fits in VRAM.
- • A new implementation allows developers to offload the majority of Qwen3.8-Flash-Next's KV cache to system RAM with minimal decode slowdown.
- • The setup achieved a 1 million token context length on vLLM using three NVIDIA 3090 GPUs.
- • The model architecture uses 12 layers with a KV cache and 36 gated delta-net layers with fixed-size recurrent states to manage memory.
- • The QSA mechanism limits the amount of KV cache data read during each decode step by selecting a maximum of 2,048 positions.
Developers can run extremely long-context local models up to 1 million tokens without being bottlenecked by limited GPU VRAM.