1. Moonshot AI Releases Kimi K3 Open-Weights Model
Moonshot AI has fulfilled its July 16 promise by making the Kimi K3 model weights publicly available on Hugging Face. The 2.8-trillion-parameter Mixture-of-Experts (MoE) model, which features a 1 million token context window and native vision capabilities, is now accessible for deployment. Due to its quantization-aware training in MXFP4, the model weights download is approximately 1.4 TB, requiring significant hardware resources such as multi-node H200 or B300 configurations for execution.
- • Moonshot AI has officially released the Kimi K3 weights on Hugging Face, following the July 16 announcement.
- • The model is a 2.8-trillion-parameter MoE architecture with 896 experts and 16 active experts per token.
- • It supports a 1 million token context window and native vision capabilities.
- • The 1.4 TB download size (MXFP4 quantization) requires significant infrastructure, such as multi-node 8x H200 or 8x B300 setups.
Developers can now access and deploy the frontier-class Kimi K3 model, though its massive 1.4 TB size necessitates high-end, multi-node hardware for local operation.
2. Celeris-1 Language Model Claims 15x Faster Response Times
Celeris-1 introduces a novel approach to low-latency LLM inference by employing a diffusion-based architecture. This design allows the general-purpose model to deliver near-GPT-5 level intelligence while operating at speeds up to 15 times faster than traditional autoregressive models. With a p50 response latency of 157ms and a throughput of 1,280 tokens per second, Celeris-1 is highly suited for real-time conversational interfaces and fast agentic loops.
- • Celeris-1 is a general-purpose language model claiming near-GPT-5 level intelligence.
- • The model features 15x faster response times compared to standard models, utilizing a new diffusion-based inference architecture.
- • It achieves a p50 response latency of 157ms.
- • Celeris-1 supports an ultra-high throughput of 1,280 tokens per second.
Developers can build highly interactive, real-time applications using a model that supports a throughput of 1,280 tokens per second.
3. NVIDIA Expands SANA Video Line with 2.0 Release
Building on the SANA-WM world model released in May 2026, NVIDIA has introduced SANA-Video 2.0. While the earlier SANA-WM focused on long-horizon generation using Gated DeltaNet, the new 2.0 release shifts to a linear attention and periodic softmax architecture. These 5B and 14B parameter models are designed to maintain high-resolution 720p output on single-GPU hardware, providing developers with a new set of tools for long-form video generation.
- • NVIDIA released SANA-Video 2.0, featuring 5B and 14B parameter models.
- • The new architecture uses linear attention and periodic softmax layers, distinct from the Gated DeltaNet used in the earlier SANA-WM.
- • The models support 720p video generation on a single GPU.
- • This release expands the SANA ecosystem following the May 2026 launch of the SANA-WM world model.
This release provides developers with larger, more capable models in the SANA family, offering an alternative architecture for high-resolution video generation compared to the previously released SANA-WM.
4. Baseten Launches Optimized 'Fast' API for GLM-5.2
Building on the existing availability of GLM-5.2, Baseten has deployed a new 'Fast' API variant specifically optimized for latency-sensitive coding and agent tasks. This update doubles the model's performance compared to its initial launch-day speeds, now achieving peak throughputs of 280 tokens per second and an average of 100 tokens per second. The company also announced plans to further improve performance through upcoming speculative decoding updates.
- • Baseten's new 'Fast' API for GLM-5.2 delivers peak speeds of 280 tokens per second.
- • The update doubles the performance of the model compared to its initial launch-day API release.
- • The variant is specifically tuned to reduce latency for coding assistants and agent-based workflows.
- • Baseten plans to integrate further performance gains via upcoming speculative decoding algorithm updates.
This optimization provides developers with a faster inference path for GLM-5.2, addressing latency requirements for interactive coding assistants and agents.
5. Feyn Open-Sources FeyNoBg Model and NoBg Training Library
Feyn has introduced a powerful open-source solution for image preprocessing with the release of FeyNoBg and its accompanying NoBg Python library. FeyNoBg is an automatic background removal model that delivers state-of-the-art performance, outperforming existing models on several benchmarks. To facilitate custom training, the open-source NoBg library provides a unified Python interface that handles the entire pipeline—from preprocessing to evaluation—currently focusing on the BiRefNet architecture. This allows developers to easily integrate high-quality image matting directly into their local workflows.
- • Feyn released FeyNoBg, an automatic background removal model trained on 26.1K diverse examples across 10 datasets.
- • The model achieves top scores on four out of eight benchmarks and performs within 2% of the leader on the remaining four.
- • Feyn open-sourced NoBg, a Python library designed to preprocess, train, and evaluate image matting models.
- • The training process for FeyNoBg involved expanding the third stage of the BiRefNet feature extractor from 18 to 24 blocks while retaining pre-trained weights.
- • NoBg currently supports the BiRefNet architecture, with plans to add more architectures in the future.
Developers can self-host a state-of-the-art background removal model and easily train custom image matting pipelines.
6. ENCFORGE Ransomware Campaign Targets Langflow Instances
Building on previous disclosures of critical vulnerabilities in Langflow, a new threat has emerged. The threat actor group JADEPUFFER is actively exploiting CVE-2025-3248—a missing-authentication flaw—to deploy the ENCFORGE ransomware. This campaign, observed in July 2026, specifically targets and destroys PyTorch, TensorFlow, and GGUF model files. With CISA now adding multiple Langflow vulnerabilities to its Known Exploited Vulnerabilities (KEV) catalog, the risk to exposed instances has escalated from unauthorized access to destructive ransomware attacks.
- • Threat group JADEPUFFER launched campaigns on July 1 and July 20 targeting internet-facing Langflow servers.
- • The attacks exploit CVE-2025-3248, a missing-authentication vulnerability allowing remote Python execution.
- • ENCFORGE ransomware is being used to encrypt and destroy PyTorch, TensorFlow, and GGUF model files.
- • CISA has added multiple Langflow vulnerabilities, including CVE-2025-3248, to its Known Exploited Vulnerabilities (KEV) catalog.
- • Security teams should update Langflow, restrict Docker socket access, and maintain immutable backups of model weights.
The threat to Langflow deployments has escalated from general vulnerability exposure to active, destructive ransomware attacks, necessitating immediate patching and enhanced security for model artifacts.
7. Shared Claude Chats and Artifacts Exposed in Search Engine Indexes
A security exposure was identified in Anthropic Claude's 'share chat' feature, which allows users to generate public links to conversations and interactive Artifacts. Because Anthropic relied on robots.txt rather than implementing 'noindex' tags on shared pages, major search engines like Google and Bing indexed these URLs, making private user conversations and business documents publicly searchable. While there is no evidence of unauthorized account access, the incident highlights the limitations of robots.txt for privacy and underscores the need for organizations to treat shared AI links with the same governance as other collaboration software.
- • Shared Claude conversations and interactive Artifacts were discovered to be indexed and searchable on Google and Bing.
- • The exposure occurred because Anthropic relied on a robots.txt file, which search engines do not respect as a directive to prevent indexing if a page is linked elsewhere.
- • A review confirmed that the shared Claude chat pages lacked the 'noindex' meta tag recommended by search engines to prevent indexing.
- • Exposed content included sensitive user data such as political advice, legal ethics, and business documents.
- • Anthropic has reportedly begun removing or making the indexed search results harder to find, but recommends that enterprises audit shared content.
Developers and teams must audit their shared Claude links and treat AI collaboration platforms with strict data governance to prevent sensitive data exposure.
8. Moonshot AI Open-Sources AgentENV for Scalable Agent Reinforcement Learning
AgentENV is a distributed platform designed for running agentic reinforcement learning environments at scale, which was used to train Moonshot's 2.8-trillion-parameter Kimi K3 model. By leveraging Firecracker microVMs, AgentENV achieves kernel-level isolation without sacrificing the speed of container-based sandboxes. The platform supports advanced state management, including pausing in under 100 ms and booting or resuming from snapshots in under 50 ms. It also allows a running sandbox to fork into up to 16 independent child sandboxes on a single node, enabling efficient parallel rollouts from a single task state.
- • Moonshot AI and kvcache-ai open-sourced AgentENV under the MIT license.
- • The platform uses Firecracker microVMs to provide kernel-level isolation for agent sandboxes.
- • It supports snapshotting, pausing, resuming, and forking, with boot and resume times under 50 ms.
- • A single running sandbox can fork into up to 16 independent child sandboxes on the same node for parallel rollouts.
- • AgentENV provides an E2B-compatible HTTP API, allowing existing E2B Python and TypeScript SDKs to work without code changes.
Developers can run secure, isolated agent environments with sub-50ms boot times and native E2B SDK compatibility.
9. Perplexity Launches Official CLI for Search API
Building on the previously released Agent API and Search as Code SDK, Perplexity has introduced pplx, a single-binary command-line tool. This CLI provides a direct interface for developers to execute live web searches and fetch cleaned page text, enabling easier integration of Perplexity's search stack into terminal-based coding agents and automated pipelines.
- • Perplexity released pplx, an official single-binary command-line client for its Search API.
- • The tool is designed for humans and coding agents, offering search and fetch commands.
- • It outputs structured data in JSON format, using standard exit codes.
- • Installation is via a shell script that installs a single binary to ~/.local/bin on macOS and Linux.
- • The tool includes token budgeting features to truncate long strings and supports non-TTY environments.
Developers can now more easily equip their coding agents and terminal workflows with real-time web search and clean markdown-like text extraction using an official CLI.
10. OpenRouter Introduces Classifiers Beta for Inference Tagging
OpenRouter's new Classifiers feature provides developers with granular observability over their LLM API usage. Currently in beta, Classifiers allow teams to attach metadata tags to their inference requests. This enables developers to categorize API calls by specific dimensions—such as task type, department, or agent complexity—making it easier to track costs, analyze performance, and debug multi-agent systems.
- • OpenRouter launched Classifiers, a new feature currently in beta.
- • The feature allows developers to tag and categorize inference activity within their workspaces.
- • Developers can organize API calls by task type, department, agent complexity, or custom criteria.
Developers can better monitor, analyze, and allocate LLM costs by tagging API calls by task, department, or agent complexity.
11. Ninfer Inference Engine Performance Boosted to 720 t/s on RTX 5090
Building on its initial release, the open-source Ninfer inference engine has achieved a significant performance increase, now delivering up to 720 tokens per second on the NVIDIA RTX 5090. This update improves upon the 542 t/s benchmark reported on July 20, further optimizing the engine for Qwen3.6 27B and 35B models with full 250k context windows. The engine remains available on GitHub with support for both Linux and Windows.
- • Ninfer performance on RTX 5090 has increased to 550–720 tokens per second.
- • This update improves upon the 542 t/s benchmark previously reported on July 20.
- • The engine continues to support Qwen3.6 27B and 35B models with 250k context windows.
- • The software remains open-source and is available on GitHub with support for Linux and Windows.
The performance gain demonstrates rapid optimization of local inference engines on flagship consumer hardware, enabling even faster response times for high-context models.