1. OpenAI Launches Agents API in Public Beta with Managed Codex Harness
OpenAI's new Agents API handles the heavy lifting of agent orchestration, including automatic context compaction for long sessions, efficient tool use, and subagent delegation. By providing a unified interface to run code, edit files, and connect to MCP servers, the API reduces the boilerplate needed to build reliable agents. Pricing is tied directly to the underlying model's API rates, alongside additional costs for hosted sandboxes and tools.
- • The Agents API provides developers with the same Codex harness and infrastructure used internally by OpenAI.
- • Developers can run agent compute in an OpenAI-managed sandbox, their own infrastructure, or nine partner sandboxes.
- • Partner sandboxes include Vercel, E2B, Modal, Cloudflare, Daytona, DigitalOcean, Blaxel, Runloop, and Oracle.
- • The API is structured around four core concepts: Agent, Environment, Session, and Events and items.
- • Data residency is currently restricted to the US, and Zero Data Retention (ZDR) is not supported.
It allows developers to build and deploy durable cloud agents using OpenAI's managed infrastructure and partner sandboxes.
2. DeepSeek Launches Updated DeepSeek-V4.1-Flash Model
Building on the original DeepSeek-V4 Flash model released in May 2026, the new V4.1-Flash iteration introduces a Causal Encoder-Decoder design and advanced Compressed Sparse Attention 2 (CSA2) to further optimize memory efficiency. This update improves upon the original architecture by utilizing FP4 quantization to reduce the KV cache footprint to 890 bytes per token, enabling superior performance on agentic and coding tasks compared to the previous V4 Pro series.
- • DeepSeek-V4.1-Flash is a 552B parameter multimodal Mixture-of-Experts model.
- • Features a 1M-token context window with 8B active parameters during prefill and 16B during decode.
- • Implements Compressed Sparse Attention 2 (CSA2) and FP4 quantization for improved KV cache efficiency.
- • Outperforms the previous DeepSeek V4 Pro 0813 on agentic, coding, and cyber benchmarks.
- • Available under an MIT license via vLLM, SGLang, Transformers, and public API.
This release provides an upgraded, more efficient multimodal model that significantly lowers memory overhead for long-context applications compared to the initial V4 Flash release.
3. Cognition Launches SWE-2 Coding Model with 50% Score on FrontierCode
Cognition's SWE-2 is built on the Kimi K3 base and has undergone extensive reinforcement learning for agentic coding. By training all reasoning-effort levels in a single run, the model optimizes the cost-performance frontier, allowing SWE-2 medium to outperform SWE-1.7 while using 58% fewer turns and costing 81% less on average. While there are no public weights or standalone API endpoints, developers can access the model immediately through the Devin suite.
- • SWE-2 is a 2.8T parameter Mixture-of-Experts model with 104B active parameters per token, post-trained from Kimi K3.
- • The model achieves a score of 50.0 on the FrontierCode 1.1 benchmark, matching Claude Fable 5.1 at a 64% lower cost.
- • It is available starting today in Devin Desktop, CLI, Devin Web, and Fusion, with no public weights or per-token API pricing.
- • SWE-2 utilizes NVFP4 and FP8 kernels for inference, alongside a prefill delayer that improves throughput by 10% to 20%.
- • The model was trained using a reinforcement learning algorithm that optimizes all reasoning-effort levels in a single run.
Developers using Devin get access to a significantly cheaper and more efficient agentic coding model that optimizes the cost-performance frontier.
4. AI Sage Releases GigaChat-3.5-Reasoning with Gated DeltaNet
GigaChat-3.5-Reasoning leverages Gated DeltaNet to achieve high efficiency in long-context scenarios. Developed using CISPO to train domain-specific experts before distilling them into a single model, it offers a powerful open-weights alternative for math, coding, and reasoning tasks. The MIT-licensed weights are now available on Hugging Face.
- • GigaChat-3.5-Reasoning is a 432B-A28B Mixture-of-Experts model featuring Gated DeltaNet.
- • The model was trained by distilling domain experts in code, math, and general knowledge via on-policy distillation.
- • It performs similarly to DeepSeek V4 Flash Preview while utilizing 37% fewer tokens in its reasoning traces.
- • Model weights are available on Hugging Face under an MIT license.
- • The model is also accessible via a web interface on giga.chat.
Developers can self-host a highly efficient reasoning model that matches frontier performance while using significantly fewer tokens in its reasoning traces.
5. CyberTiel 35B-A3B Released for Offensive Security and Agentic Coding
CyberTiel targets real-world software engineering problems, with its creator claiming it beats Opus 4.6 medium on codebase issues in a fraction of the time required by Qwen3.8-27b. By sacrificing general world knowledge, the model maximizes its utility for specialized coding and security tasks. Developers can run the model locally using the newly released GGUF and MLX formats.
- • CyberTiel is an uncensored 35B-A3B coder model based on the TielCoder base.
- • The model is designed specifically for offensive security research and agentic software engineering.
- • It was created using quantization with an improved imatrix trained on a curated cybersecurity and software engineering corpus.
- • GGUF and MLX versions of the model are available on Hugging Face for local deployment.
- • The model prioritizes coding ability over general world knowledge and is developed by an anonymous researcher.
It provides security-focused developers with a highly specialized, uncensored local model that outperforms larger models on real-world codebase issues.
6. OUI-1 Model Released for Generating Bespoke UI Elements
OUI-1 introduces an efficient method for generating user interface elements by training on OpenUI-Lang rather than verbose standard web formats. This DSL-centric approach significantly cuts down on the system prompt overhead typically required for UI formatting. However, developers looking to run the model locally will face hurdles, as the underlying DiffusionGemma architecture lacks llama.cpp support.
- • OUI-1 is a model fine-tuned on DiffusionGemma and uploaded to Hugging Face.
- • The model uses OpenUI-Lang, a custom domain-specific language, instead of HTML, Markdown, or React.
- • Fine-tuning on a DSL is designed to reduce context window overhead compared to system prompts.
- • DiffusionGemma is not currently supported by llama.cpp, limiting local execution via tools like Ollama.
It offers a specialized approach to UI generation that reduces token usage, though local deployment is currently limited by llama.cpp compatibility.
7. LandingAI Releases Agentic Document Extraction Gen2 with DPT-3 Models
LandingAI's ADE Gen2 represents a major shift in document processing by moving away from flat chunking to hierarchical tree structures. This allows for precise data extraction with verifiable grounding. The release is generally available and supports flexible deployment options, including AWS, Azure, Google Cloud, Snowflake, and on-premises, with projected cost savings of 25% to 80% on mixed workloads.
- • ADE Gen2 is built on the new DPT-3 model family, featuring DPT-3 Pro and DPT-3 Verity.
- • The system treats documents as tree structures and provides atomic grounding to trace data to specific words or lines.
- • DPT-3 Verity offers deterministic transcription with word-level confidence, while DPT-3 Pro handles complex layouts and handwriting.
- • Pricing has shifted from a flat 3-credit-per-page model to a per-page rate plus an output character rate.
- • Gen1 client code is incompatible with Gen2 endpoints, requiring a migration for existing users.
Developers get highly precise document parsing with line-level grounding, though they must migrate existing client code to the new incompatible endpoints.
8. Redis Launches LangCache Managed Semantic Caching Service
Redis LangCache operates as a two-call loop, searching the semantic cache before invoking the LLM and storing the response afterward. By matching the intent of a prompt rather than its exact text, the service maximizes cache hit rates for dynamic user inputs. Redis ensures that customer data remains on their own servers and is never used for model training.
- • Redis LangCache is a managed semantic caching service available in public preview on Redis Cloud.
- • The service is accessible via a REST API with dedicated Python and JavaScript SDKs.
- • It matches prompts based on semantic meaning, returning cache hits up to 15x faster than standard LLM queries.
- • Features include configurable similarity thresholds, TTLs, eviction policies, and adaptive precision controls.
- • Early customer Mangoes.ai reported a 70% cache hit rate and a 70% reduction in LLM spending.
It reduces LLM API costs and response latency through managed semantic caching on Redis Cloud.
9. Sentry Introduces Agent Tracing Tool for Debugging LLM Workflows
Sentry's new agent tracing capabilities address the observability challenges of non-deterministic AI workflows. By mapping out agent execution steps, developers can easily trace malformed outputs back to specific tool calls or prompt failures. The tool was showcased in a live workshop demonstrating debugging across diverse environments like Slack and GitHub Actions.
- • Sentry's agent tracing tool diagnoses issues like bad tool calls, malformed output, and unaccounted token spend.
- • The tool allows developers to identify exactly where agents break and the associated operational costs.
- • Sentry demonstrated the tool on live agents, including a Slack agent, an e-commerce chatbot, and a GitHub Action.
- • Developers can integrate their own custom agents into the Sentry platform for monitoring and analysis.
It gives developers deep visibility into agent execution paths to pinpoint failures and optimize token spend.
10. NVIDIA Releases SoL-Pi Extension to Optimize Pi Coding Agent
NVIDIA's SoL-Pi extension provides a non-invasive way to optimize the Pi coding agent's performance. By implementing features like the Evidence-Preserving Reducer to compact diagnostic logs and Online Context Compact to manage window pressure, the extension minimizes unnecessary token exchange. Because all mechanisms are opt-in and disabled by default, developers can selectively apply optimizations to match their specific workflows.
- • SoL-Pi is a standalone extension designed to improve the efficiency of the Pi coding agent.
- • The extension installs on top of an unmodified Pi release using public APIs without patching source code.
- • It includes four opt-in mechanisms: Action Fusion, ObservationPack, Evidence-Preserving Reducer, and Online Context Compact.
- • Action Fusion runs validation commands in the same tool call as edit or write operations.
- • ObservationPack converts large text results into stable handles with paged recall to reduce token traffic.
It allows developers using the Pi agent to significantly lower inference costs and latency through opt-in optimization mechanisms.
11. LangSmith Connections Launches for Managed Agent Credential Management
Managing credentials in agentic workflows has historically been a security bottleneck. LangSmith Connections addresses this by providing a structured way to handle OAuth and shared API keys. This allows agents to perform authenticated tasks, such as filing tickets or querying external databases, while maintaining strict security boundaries and audit trails.
- • LangSmith Connections is a managed credential management system designed for AI agents.
- • The system supports agent-owned credentials for shared tasks like web searches.
- • It supports user-owned OAuth setups, allowing agents to act securely on behalf of individual users.
- • The system aims to enhance both the security and functional capabilities of deployed agents.
It solves the security challenge of managing API keys and OAuth tokens for autonomous agents executing web searches or filing tickets.
12. ZeroModels Releases Keras 3 Pretrained Models for Multi-Backend Inference
ZeroModels provides a streamlined alternative for developers integrating multimodal features into their applications. By leveraging Keras 3, the collection allows models to run seamlessly on JAX, PyTorch, or TensorFlow depending on the deployment environment. The lack of heavy runtime dependencies like transformers makes it highly suitable for resource-constrained or highly optimized production pipelines.
- • ZeroModels is a collection of pretrained models built entirely in Keras 3.
- • Supported tasks include image classification, object detection, segmentation, depth estimation, and speech recognition.
- • The codebase is fully compatible with JAX, PyTorch, and TensorFlow backends.
- • The collection eliminates the need for transformers or torch libraries at runtime.
It allows developers to deploy lightweight vision and speech models across different backends without heavy runtime library overhead.
13. Cherenkov Inference Engine Optimizes Local MoE Models on Apple Silicon
Cherenkov addresses the memory limitations of running massive Mixture-of-Experts models on local hardware. By predicting which experts will be needed and streaming them from the SSD on the fly, the engine avoids loading the entire model into RAM. This predictive streaming approach enables memory-constrained devices, like a 32GB MacBook Air, to run models that would otherwise exceed their physical memory limits.
- • Cherenkov is an inference engine designed specifically for Apple Silicon.
- • The engine uses predictive expert streaming to maintain a bounded working set of experts in unified memory.
- • It utilizes a one-layer lookahead to predict required experts and initiate SSD reads in advance.
- • The system can fall back to smaller Q3 or Q2 quantizations if SSD read times are insufficient.
- • It achieved 8-22 tokens per second running Qwen3.8-Flash-Next on a 32GB M4 MacBook Air.
It allows developers to run large Mixture-of-Experts models locally on consumer-grade Macs by streaming experts from SSD on demand.