1. Alibaba Launches Hosted Qwen3.8-Omni-Flash API
Following the August release of the open-weight Qwen3.8-Flash-Next model, Alibaba has now launched Qwen3.8-Omni-Flash as a hosted API service. This new omni-modal version supports text, image, audio, and video inputs with a 1 million token context window, available via QwenCloud and Alibaba Cloud Model Studio. To support developers building on this model, the team has released Qwen-MM-Plugins under an Apache-2.0 license to facilitate agentic harnesses.
- • Qwen3.8-Omni-Flash is a hosted API version of the Qwen3.8-Flash-Next architecture.
- • The service supports text, image, audio, and video inputs with a 1 million token context window.
- • Pricing is set at $0.15 per 1 million input tokens and $0.47 per 1 million output tokens.
- • The release includes Qwen-MM-Plugins (Apache-2.0) to support agentic harnesses.
- • The model improves OmniVideoBench accuracy to 67.8 while reducing token usage by 45.7%.
Developers can now access the Qwen3.8-Flash-Next architecture as a managed service for multimodal applications, including long-form video and audio processing, without needing to host the model themselves.
2. xAI Releases Grok Imagine Image 2.0 with Multi-Reference Support
xAI has released Grok Imagine Image 2.0, a text-to-image and editing model that ranks fourth on the Artificial Analysis Text to Image Leaderboard, making it the highest-ranked model outside of OpenAI. The model supports text-to-image generation, image editing, and multi-reference generations using up to five input images. It is available via the xAI API, the Grok Imagine website, Grok apps, fal, and Replicate. Significant performance gains were observed in knowledge, text rendering, and lighting capabilities compared to the previous generation.
- • xAI released Grok Imagine Image 2.0, which ranks fourth on the Artificial Analysis Text to Image Leaderboard.
- • The model supports text-to-image generation, image editing, and multi-reference generations using up to five input images.
- • It is available via the xAI API, the Grok Imagine website, Grok apps, fal, and Replicate.
- • The model is positioned on the Pareto frontier for quality versus price on the Artificial Analysis leaderboard.
- • Significant performance gains were observed in knowledge, text rendering, and lighting capabilities.
Developers can integrate a highly competitive, cost-effective image generation and editing model that supports multi-reference generation using up to five input images.
3. SpaceXAI Releases Grok Voice Transcribe 2.0 with Low-Latency Streaming
SpaceXAI has released Grok Voice Transcribe 2.0, a new speech-to-text model succeeding Grok Voice Transcribe 1.0. The model achieves a 2.7% Word Error Rate (WER) at 0.49 seconds after the end of speech for streaming Final Transcripts, and a 3.4% WER for First Partial Transcripts. In non-streaming tasks, the model scores 2.3% on AA-WER, ranking 5th out of 59 models with a speed factor of approximately 160x. While more accurate, it is slower than competing models such as Muse Voice Transcribe, ElevenLabs Scribe v2 Realtime, and Cartesia Ink-2.
- • SpaceXAI released Grok Voice Transcribe 2.0, succeeding the previous 1.0 version.
- • The model achieves a 2.7% Word Error Rate (WER) at 0.49 seconds after the end of speech for streaming Final Transcripts.
- • Streaming transcription is priced at $0.20 per hour ($3.33 per 1,000 minutes), while non-streaming is $0.10 per hour ($1.67 per 1,000 minutes).
- • In non-streaming tasks, the model scores 2.3% on AA-WER, ranking 5th out of 59 models with a speed factor of approximately 160x.
- • The streaming price is lower than Cartesia Ink-2 ($4.00) and approximately half the cost of ElevenLabs Scribe v2 Realtime ($6.50).
Developers can implement highly accurate, low-latency streaming voice transcription in their applications at a competitive price of $0.20 per hour.
4. Jina AI Releases jina-ocr-v1 for Low-Budget GPUs
Jina AI has released jina-ocr-v1, an end-to-end visual document parser designed for low-budget GPUs such as the NVIDIA L4. The model contains 3.4 billion total parameters with approximately 570 million active parameters per token, built on the DeepSeek-OCR backbone and utilizing a DeepSeek-3B-MoE decoder. It includes a FastMTP speculative decoding head that enables lossless, near-deterministic output generation. The parser outputs content in Markdown, with tables formatted in HTML and formulas in LaTeX, achieving a throughput of 2.57 pages per second on a single NVIDIA A100 40GB GPU.
- • Jina AI released jina-ocr-v1, an end-to-end visual document parser designed for low-budget GPUs.
- • The model contains 3.4 billion total parameters with approximately 570 million active parameters per token.
- • It is built on the DeepSeek-OCR backbone and utilizes a DeepSeek-3B-MoE decoder with a FastMTP speculative decoding head.
- • The model achieved scores of 91.14 on OmniDocBench v1.6 and 83.4 on olmOCR-Bench.
- • The model is released under a CC BY-NC 4.0 license, requiring commercial licensing from Jina AI.
- • The parser outputs content in Markdown, with tables formatted in HTML and formulas in LaTeX.
Developers can self-host a highly accurate document parser that outputs structured Markdown, HTML tables, and LaTeX formulas at high throughput.
5. Realtime-Venus Open-Sources Full-Duplex Audio-Visual Models
The Realtime-Venus repository has released two checkpoints: Realtime-Venus-Omni and Realtime-Venus-Audio. Realtime-Venus-Omni is a 9B audio-visual interaction model adapted from MiniCPM-o 4.5 that supports proactive interaction, semantic interruption handling, and training-free long-video memory. Realtime-Venus-Audio is an audio-focused checkpoint designed for audio understanding and audio-driven conversation. The system features native full-duplex conversation, allowing it to perceive input while simultaneously speaking, and uses in-stream delegate requests to handle external tool integrations without blocking ongoing conversations.
- • The Realtime-Venus repository hosts Realtime-Venus-Omni (9B audio-visual model) and Realtime-Venus-Audio (audio-focused model).
- • The system features native full-duplex conversation, allowing it to perceive input while simultaneously speaking.
- • The Omni-Proactive interaction feature enables the model to initiate responses to events without waiting for a user prompt.
- • It uses in-stream delegate requests to handle external tool integrations without blocking ongoing conversations.
- • The system generates both text and native speech output using bundled Token2wav resources and a reference voice.
Developers can build highly interactive voice and video agents capable of semantic interruption handling and in-stream tool integration.
6. Laya 421M Non-Autoregressive Decision Model Released
Developer Laya has released a 421M-parameter non-autoregressive decision model designed for tasks including intent routing, fact-checking, moderation consensus, prompt guardrails, rubric scoring, and multi-turn conversation trajectories. The model architecture pairs a bidirectional ModernBERT-large encoder with a scratch Transformer head to resolve typed schemas, achieving an extremely fast forward pass time of approximately 35 milliseconds. It was trained on a single RTX 6000 Pro GPU with 96 GB of VRAM using over 25,000 human-annotated real-world examples, optimized via a policy-gradient reinforcement learning approach.
- • Laya is a 421M-parameter non-autoregressive decision model for intent routing, fact-checking, moderation, and guardrails.
- • The architecture pairs a bidirectional ModernBERT-large encoder with a scratch Transformer head to resolve typed schemas.
- • The model achieves an extremely fast forward pass time of approximately 35 milliseconds.
- • It was trained on a single RTX 6000 Pro GPU using over 25,000 human-annotated real-world examples.
- • Laya is available for testing via a dedicated Hugging Face space and a GitHub repository.
Developers can run a highly specialized, ultra-fast routing and moderation model locally with a forward pass time of just 35 milliseconds.
7. Confucius4-R2T2 Open-Source Streaming ASR Model Released
A new real-time streaming automatic speech recognition (ASR) model called Confucius4-R2T2 has been open-sourced. The model is a 1.7 billion parameter system specifically designed for voice agents. It is available for access via GitHub and Hugging Face, providing developers with a self-hostable option for real-time speech-to-text tasks.
- • Confucius4-R2T2 is an open-source, real-time streaming automatic speech recognition (ASR) model.
- • The model features 1.7 billion parameters and is specifically designed for voice agents.
- • It is available for access via GitHub and Hugging Face.
Developers can self-host a low-latency, voice-agent-optimized speech-to-text model directly from GitHub or Hugging Face.
8. Claude Code Adds Native Windows Support and AGENTS.md Integration
Anthropic has expanded the capabilities of Claude Code, building on the recently released plugin evaluation workflow. The latest update adds native Windows installation support, PDF reading, and the ability to read AGENTS.md files for project configuration. It also introduces dynamic workflows for multi-agent orchestration and new proxy configuration options, alongside expanded model support including Claude Opus 4.8, 4.5, and Sonnet 4.6.
- • Claude Code now supports native Windows installations and PDF file reading.
- • The tool reads AGENTS.md for project-specific instructions.
- • New dynamic workflows enable complex multi-agent task orchestration.
- • Added support for Claude Opus 4.8, 4.5, Sonnet 4.6, and Amazon Bedrock via Mantle.
- • New environment variable CLAUDE_GATEWAY_PROXY_IS_EGRESS_BOUNDARY=1 supports forward proxies.
Developers gain improved cross-platform accessibility and project-level configuration control, further maturing the Claude Code agentic workflow.
9. ZCode Coding Agent Found Silently Uploading Git History to Aliyun OSS
A security investigation into ZCode, Zhipu's AI coding desktop application, has revealed that the software silently packages and uploads user workspaces to Aliyun OSS. The captured data includes the entire .git history, LFS asset cache, reflogs, and global configurations, with the .git directory accounting for approximately 86.6% of the payload. The background capture and upload process is triggered unconditionally upon startup as long as the user is logged in, completely ignoring UI settings like 'Optimize Experience' and 'Repo Snapshot Indexing.' The uploaded archives are encrypted using envelope encryption with a public key, meaning users cannot access or decrypt their own uploaded files.
- • ZCode silently packages, encrypts, and uploads user workspaces, including .git history, LFS asset caches, and reflogs, to Aliyun OSS.
- • The background capture runs unconditionally at startup when logged in, ignoring UI toggles like 'Optimize Experience' and 'Repo Snapshot Indexing.'
- • The uploaded archives use envelope encryption with a public key, leaving the private key exclusively in Z.ai's control.
- • The .git directory accounts for roughly 86.6% of the payload, risking exposure of historical API keys and deleted sensitive configurations.
- • Users can block the upload process by using filesystem-level immutability flags (chflags on macOS or chattr on Linux) on the '~/.zcode/v2/checkpoints' directory.
Developers using ZCode must take immediate action to block background uploads to prevent the exposure of historical API keys, deleted configurations, and proprietary code.
10. MiniMax Open-Sources Terminal Version of MiniMax Code
MiniMax has open-sourced the terminal version of MiniMax Code on GitHub under the MIT license. The release is a 0.4.12 source preview and does not include the source for the desktop application. The repository features an interactive TUI, headless execution, code editing, shell commands, diffs, test verification, permission controls, and sandboxing. It supports subagents, plugins, skills, Model Context Protocol (MCP), and Bring Your Own Key (BYOK) with OpenAI- and Anthropic-compatible providers.
- • MiniMax open-sourced the terminal version of MiniMax Code (v0.4.12 source preview) on GitHub under the MIT license.
- • The tool features an interactive TUI, headless execution, code editing, shell commands, diffs, and sandboxing.
- • It supports subagents, plugins, skills, Model Context Protocol (MCP), and Bring Your Own Key (BYOK) for OpenAI- and Anthropic-compatible providers.
- • The release does not include the source code for the desktop application.
- • The repository notes that matching version numbers do not guarantee identical build provenance between the published package and source checkout.
Developers can adopt a highly customizable, MIT-licensed terminal coding agent that supports subagents, plugins, and custom LLM backends.
11. Notion Adds Skills API to Developer Platform for Agent Instruction Management
Following the May 2026 launch of its developer platform for AI agents, Notion has introduced a Skills API. This new tool allows teams to collaboratively edit and distribute agent instructions, integrating with GitHub sync or Vercel's skills installer to streamline deployment across agent environments.
- • Expands the Notion developer platform with a new Skills API.
- • Enables collaborative editing and distribution of agent instructions.
- • Supports deployment via GitHub sync and Vercel's skills installer.
This update provides the specific infrastructure needed to manage and share agent instructions at scale, evolving the platform from a workspace integration hub to a collaborative agent development environment.
12. NVIDIA Open-Sources SoL-Pi Agent Harness on Hugging Face
NVIDIA has expanded its SoL-Pi project by releasing the full agent harness as an open-source tool on Hugging Face. This follows the earlier release of the SoL-Pi extension, which was designed specifically to optimize the Pi coding agent. The new harness leverages recursive auto-research loops to reduce token traffic by up to 49%, providing a broader framework for developers to cut API costs across various agentic workflows.
- • NVIDIA has open-sourced the SoL-Pi agent harness on Hugging Face.
- • This release expands on the previously announced SoL-Pi extension for the Pi coding agent.
- • The harness uses recursive auto-research loops to reduce token traffic by 44.7% to 49.0%.
- • Developers can reduce API costs by approximately one-third while maintaining agent performance.
Developers can now access the full SoL-Pi harness to optimize token usage and reduce API costs in a wider range of agentic applications beyond the initial Pi coding agent extension.
13. Guide Ranks Top Open-Source Agent Harnesses for Local LLMs
A guide published on September 18, 2026, ranks 11 open-source agent harnesses for local LLMs based on licensing, documentation, maintenance, and safety. The guide highlights OpenCode, Goose, Cline, OpenHands, and Aider, among others. It outlines general best practices, such as setting context windows to at least 64,000 tokens and ensuring models support tool calling. It also details specific hardware requirements, noting that OpenHands recommends using Qwen3.6-35B-A3B and requires at least 24GB of VRAM or 64GB of unified memory on Apple Silicon.
- • The guide ranks 11 open-source agent harnesses, including OpenCode, Goose, Cline, OpenHands, and Aider.
- • General best practices recommend setting context windows to at least 64,000 tokens and ensuring models support tool calling.
- • OpenHands recommends using Qwen3.6-35B-A3B and requires at least 24GB of VRAM or 64GB of unified memory on Apple Silicon.
- • Goose supports runtimes including Ollama, LM Studio, and Docker Model Runner under the Linux Foundation's Agentic AI Foundation.
- • Aider manages weak tool calling by using repository maps and specific text-based edit formats.
Developers can quickly evaluate and select the right local agent framework based on concrete hardware constraints and licensing terms.
14. Empirical Study Identifies Optimal Harness Designs for Coding Agents
A new empirical study evaluated coding agents across 176 different settings using SWE-Bench Verified and Terminal-Bench 2.1 to analyze planning, action space, and context management. The study found that context management is the most critical factor for performance under tight computational budgets, with rule-based elision staged before LLM-based summarization proving to be the most efficient strategy. Additionally, planning acts as an accuracy scaffold for weaker models and a cost-saving mechanism for stronger models, while predefined tools improve performance for models with weaker bash proficiency.
- • Researchers evaluated 176 settings across four models using SWE-Bench Verified and Terminal-Bench 2.1.
- • Context management was identified as the most critical factor for performance under tight computational budgets.
- • Staging rule-based elision before LLM-based summarization proved to be the most efficient context-management strategy.
- • Planning acts as an accuracy scaffold for weaker models and a cost-saving mechanism for stronger models.
- • Predefined tools improve performance for models with weaker bash proficiency, whereas bash-capable models can operate effectively using only a bash interface.
Developers building custom coding agents can optimize performance and reduce API costs through structured context management and adaptive planning.
15. Run 95.5 GiB Qwen3.8-Flash-Next on a 64GB Mac via Expert Streaming
Developers can now run the 95.5 GiB Qwen3.8-Flash-Next model on a Mac with 64 GB of RAM by keeping routed experts on an SSD. This setup requires a custom fork of llama.cpp to support expert streaming and the nitinpanj/qwen38-flash-next-v3 checkpoint. On an M5 Pro with 64 GB of RAM, the model achieves prompt processing speeds of approximately 367 tokens per second at 4k context, and generation speeds reach 27.6 tokens per second with a draft head enabled. Performance is further enhanced by gather-based sparse attention and Metal MoE fusion.
- • The 95.5 GiB Qwen3.8-Flash-Next model can run on a 64GB Mac by keeping routed experts on an SSD.
- • The setup requires a custom fork of llama.cpp to support expert streaming and the nitinpanj/qwen38-flash-next-v3 checkpoint.
- • On an M5 Pro with 64 GB of RAM, generation speeds reach 27.6 tokens per second with a draft head enabled.
- • Gather-based sparse attention improved performance by 19% at 62k context and 50% at 130k context.
- • The practical ceiling for the expert cache on a 64 GB Apple Silicon machine is 36 GiB.
Developers can run massive, high-quality Mixture-of-Experts models locally on standard Apple Silicon hardware without needing expensive multi-GPU setups.