1. Tencent Releases Hy4 Preview, a 770B Parameter Open-Weight Model
Tencent has released Hy4 Preview, an open-weight large language model featuring 770 billion total parameters and 49 billion active parameters. The model supports a 1 million token context window and is available on Hugging Face, though its massive 1.56TB file size presents substantial hosting requirements. Hy4 introduces a dual-level reasoning feature, allowing developers to toggle between a default 'high' reasoning level and a 'no_think' level that disables reasoning entirely.
- • Hy4 is an open-weight text input large language model containing 770B total parameters and 49B active parameters.
- • The model supports a 1M token context window.
- • The model file size is 1.56TB on Hugging Face.
- • It features two reasoning levels: a default 'high' level and a 'no_think' level that disables reasoning.
Developers can experiment with a massive open-weight model that supports a 1M context window and allows disabling reasoning via a 'no_think' level.
2. OpenAI Details ChatGPT Work Subscription Tiers and Local Automation Features
Building on the initial launch of ChatGPT Work in July, OpenAI has detailed the product's subscription structure. The service is now split into Work Cloud, which provides a code execution environment and headless Chrome for web automation, and Work Local, which enables direct access to local files via the desktop application. Additionally, users can now deploy stateful websites using ChatGPT Sites powered by Cloudflare Workers, alongside new support for sub-agent sessions and scheduled prompt automations.
- • ChatGPT Work is now divided into Work Cloud and Work Local tiers.
- • Work Local enables direct access to local files through the desktop application.
- • Work Cloud includes a persistent filesystem, code execution, and headless Chrome for web automation.
- • Users can deploy stateful websites via ChatGPT Sites using Cloudflare Workers.
- • The platform now supports sub-agent sessions and scheduled prompt automations.
These features transform ChatGPT Work from a general agent into a specialized development environment capable of local file manipulation and automated web deployment.
3. Google AI Releases TimesFM-3 for Multivariate Time Series Forecasting
Google Research has released TimesFM-3, a 330 million parameter zero-shot foundation model designed for multivariate time series forecasting. Pretrained on over 1 trillion time points, the model supports multiple targets, past covariates, and past-future covariates without requiring task-specific fine-tuning. It utilizes a decoder-only transformer architecture with alternating causal temporal attention and full variate attention to model cross-series dependencies. While the repository code is licensed under Apache-2.0, developers should note that the model weights are restricted to non-commercial and non-production use.
- • Google Research released TimesFM-3, a 330 million parameter foundation model for multivariate time series forecasting.
- • The model was pretrained on more than 1 trillion time points of real and synthetic data.
- • It supports multiple targets, past covariates, and past-future covariates without task-specific fine-tuning.
- • The architecture uses a decoder-only transformer with alternating causal temporal attention and full variate attention.
- • TimesFM-3 weights are restricted to non-commercial and non-production use, while the repository code is licensed under Apache-2.0.
- • It ranks first among pretrained foundation models on the GIFT-Eval, fev-bench, and TIME benchmarks.
Developers can use a state-of-the-art zero-shot model for complex multivariate forecasting, though they must comply with its non-commercial license.
4. Apodex 1.1 Benchmark Results and API Pricing Detailed
Building on the August 27 announcement of the Apodex 1.1 model family, Apodex has now released detailed performance metrics and commercial pricing. The model scored 44 on the Artificial Analysis Intelligence Index and achieved an Elo of 1348 on the GDPval-AA v2 agentic benchmark. While it shows strong performance on TerminalBench v2.1 with a 70% score, it also exhibited a 78.4% hallucination rate on the AA-Omniscience benchmark. The model is now available via the Apodex first-party API at $0.30 per million input tokens and $3.00 per million output tokens.
- • Apodex 1.1 scored 44 on the Artificial Analysis Intelligence Index.
- • The model achieved an Elo of 1348 on the GDPval-AA v2 agentic benchmark.
- • TerminalBench v2.1 performance reached 70%.
- • API pricing is set at $0.30 per 1M input tokens and $3.00 per 1M output tokens.
- • The model features a 256K token context window.
- • Performance on the AA-Omniscience benchmark showed a 78.4% hallucination rate.
Developers can now evaluate the cost-to-performance trade-offs of the Apodex 1.1 model using standardized benchmark data and official pricing, allowing for informed integration into agentic workflows.
5. DeepSeek-V4-Pro-0813-NVFP4 Quantized Model Released
The DeepSeek-V4-Pro-0813-NVFP4 model has been released, providing a quantized version of DeepSeek's autoregressive Mixture-of-Experts language model. Quantized using Model Optimizer, this release is tailored for advanced reasoning, tool use, and agentic AI applications. The model is open for both commercial and non-commercial use, making it accessible for production deployments.
- • DeepSeek-V4-Pro-0813-NVFP4 is a quantized version of the autoregressive Mixture-of-Experts language model DeepSeek-V4-Pro-0813.
- • The model was quantized using Model Optimizer.
- • It is designed for advanced reasoning, agentic AI applications, tool use, and complex problem-solving.
- • The model is available for both commercial and non-commercial use.
Developers can deploy a highly optimized, quantized version of DeepSeek's reasoning model for agentic workflows and complex problem-solving.
6. DeepSeek-V4-Flash-Vision-Exp Weights Now Available on Hugging Face
Building on the August 21 announcement of the DeepSeek-V4-Flash-Vision-Exp experimental model, DeepSeek has now made the model weights available for download on Hugging Face. This release allows developers to host the vision-capable model locally, complementing the previously announced API-based access.
- • DeepSeek-V4-Flash-Vision-Exp weights are now hosted on Hugging Face.
- • This follows the initial August 21 announcement of the model's API-based vision capabilities.
- • The release enables local deployment and self-hosting of the experimental vision-capable model.
Developers can now move beyond API-only access to self-host the experimental vision model for custom multimodal applications.
7. ContextPilot-14B Released to Enhance Agent Planning and Memory
ContextPilot-14B has been released on Hugging Face as a specialized Qwen3-14B checkpoint designed to enhance agent capabilities. The model focuses on teaching agents how to plan, maintain long-term memory, and offload less useful context while continuing to reason and execute tools. This provides developers with a targeted open-weight option for building more robust autonomous workflows.
- • ContextPilot-14B is a Qwen3-14B checkpoint hosted on Hugging Face.
- • The model is designed to enhance agent capabilities.
- • It teaches agents to plan, maintain long-term memory, and offload less useful context while continuing to reason and use tools.
Developers building autonomous agents can use this specialized checkpoint to improve long-term reasoning and context management.
8. Nanbeige4.2-3B-DSpark Released for Accelerated Inference
Building on the Nanbeige4.2-3B model introduced in July, the new Nanbeige4.2-3B-DSpark variant has been released on Hugging Face. This version is specifically optimized for resource-constrained environments, achieving auto-regressive generation speeds of approximately 35 tokens per second. It is positioned as a faster alternative to the Qwen 3.5 9B model for developers with limited GPU resources.
- • Nanbeige4.2-3B-DSpark is a new, optimized version of the Nanbeige4.2-3B model released in July.
- • The model is hosted on Hugging Face and designed for resource-constrained environments.
- • It achieves auto-regressive generation speeds of approximately 35 tokens per second.
- • It is positioned as a faster alternative to the Qwen 3.5 9B model for limited GPU setups.
This release provides an optimized, high-speed inference path for the existing Nanbeige4.2-3B architecture, enabling faster performance on limited hardware.
9. Anthropic Clarifies Auto Mode Security Following New Code Execution Exploit
A security researcher has demonstrated an attack chain against Claude Code Opus 5 in Auto Mode that achieves arbitrary code execution by exploiting Python module shadowing. This finding follows Anthropic's August 14 transition to making Auto Mode the default setting for Claude Code. In response to the exploit, Anthropic clarified that Auto Mode is a convenience feature with a best-effort classifier, not a security boundary, despite previous claims regarding its safety. Developers are advised to run coding agents in isolated environments like containers or virtual machines and restrict network egress.
- • A researcher demonstrated an attack chain against Claude Code Opus 5 in Auto Mode achieving code execution with up to an 80% success rate.
- • The exploit uses Python module shadowing to trick the agent into executing malicious code from a poisoned ZIP archive.
- • Anthropic clarified that Auto Mode is a convenience feature, not a security boundary, countering previous positioning of the mode as a secure default.
- • Developers are urged to use isolated environments (containers/VMs) and restrict network egress when using coding agents.
The disclosure highlights a gap between the perceived security of default agentic modes and their actual resilience, necessitating stricter local environment controls.
10. OpenClaw 2.0 Released with Guided Setup and Rebuilt Control UI
The OpenClaw team has released OpenClaw 2.0, the largest update in the open-source project's history, incorporating over 16,000 pull requests. Designed for single-operator and team deployments, the release features a rewritten installation process that automatically detects and verifies credentials for providers like OpenAI, Anthropic, Ollama, and LM Studio. The rebuilt Control UI slashes startup time to 575 milliseconds, while new docked panels add a workspace file editor, a git-backed Changes panel, and a full-screen web terminal. Sessions and transcripts have also been migrated to SQLite storage.
- • OpenClaw 2.0 features a rewritten installation process that automatically detects existing subscriptions, API keys, and local models.
- • The rebuilt Control UI reduces startup time from 1.6 seconds to 575 milliseconds and decreases JavaScript requests from 140 to 45.
- • Sessions and transcripts have been migrated to SQLite storage, and new docked panels provide a workspace file editor, git-backed Changes panel, and web terminal.
- • Shared cloud sessions enable multi-user collaboration, though they do not serve as a security boundary or tenant isolation.
- • The release incorporates over 16,000 pull requests from 933 contributors, representing approximately 50% of all pull requests ever merged into the project.
Developers deploying single-operator or team AI gateways can leverage a faster, more integrated workspace with built-in model verification and collaborative sessions.
11. uv 0.12.8 Introduces File-Level Cache Deduplication Using BLAKE3
Astral has released uv version 0.12.8, introducing file-level deduplication in the wheel cache. By storing files under their BLAKE3 hash in a files-v0 bucket and using hardlinks to place objects in their original locations, the update significantly reduces disk usage without altering existing installation workflows. The release also optimizes cache cleanup on macOS using getattrlistbulk to read hard-link counts in batches, speeds up dependency graph construction, reduces allocations during wheel extraction, and prevents concurrent redundant downloads.
- • uv version 0.12.8 introduces file-level deduplication in the wheel cache by storing files under their BLAKE3 hash in a files-v0 bucket.
- • The deduplication feature uses hardlinks to store objects in their original locations, maintaining existing workflows.
- • Cache cleanup for the content-addressed cache is optimized for macOS using getattrlistbulk to read hard-link counts in batches.
- • Performance improvements include faster dependency graph construction, reduced allocations during wheel extraction, and prevention of concurrent redundant downloads.
- • Bug fixes address Azure Storage API compatibility, URL redaction for Azure shared access signatures, and workspace discovery.
Developers using uv will experience faster dependency graph construction, reduced disk usage, and optimized cache cleanup on macOS.
12. OpenAI Codex Memory Feature Allegedly Exfiltrates Local Chat Content
A security concern has emerged regarding OpenAI Codex's memory feature, which allegedly exfiltrates local-provider chat content to OpenAI's servers. According to reports, this data transfer occurs without providing notice to the user, raising privacy and security concerns for developers who utilize local LLM environments alongside OpenAI's developer tools.
- • OpenAI Codex's memory feature allegedly exfiltrates local-provider chat content to OpenAI.
- • The data transfer occurs without providing notice to the user.
- • This affects developers who mix local and cloud-based development tools.
Developers using local LLM providers alongside OpenAI tools should be aware of potential data leakage risks affecting their private codebases.
13. Memoryfields Introduces Portable Agent Memory File Format
Memoryfields has been introduced as a portable, low-mechanism agent memory file format. Composed of Markdown pages, optional YAML frontmatter, and an optional SQLite vector index for semantic search, the format treats memory strictly as data. This design avoids complex RAG pipelines like chunking or re-ranking, allowing agents to interact with memory using simple access patterns like bash or SQLite. Memoryfields are packaged as zip files and are transport-agnostic, supporting local storage, Amazon S3, GitHub, or HTTP. A CLI tool and management skill are available, with nomic-embed-text-v1.5 recommended for embeddings.
- • Memoryfields is a portable agent memory file format consisting of Markdown pages, optional YAML frontmatter, and an optional SQLite vector index.
- • The format treats memory as data, avoiding complex RAG processes like chunking or re-ranking.
- • Each memory page has a soft limit of approximately 8kb or 2000 tokens.
- • Memoryfields are stored as zip files but are transport-agnostic, supporting local storage, Amazon S3, GitHub, or HTTP.
- • The author provides a CLI tool and a skill for managing memoryfields, recommending the nomic-embed-text-v1.5 model for embeddings.
Developers can use Memoryfields to implement low-mechanism, transport-agnostic agent memory without complex RAG pipelines.
14. Hebbian Robotics Releases HFlow SDK for Multimodal Robotics Data
Hebbian Robotics has open-sourced HFlow, an SDK designed to build scalable robotics data pipelines. Released under the Apache-2.0 license, HFlow converts raw multimodal recordings—such as video, joint states, and actions—into standardized episodes and queryable dataset manifests. The SDK uses the MCAP container format to keep sensor streams synchronized. Pipelines are written as Python functions that can run as Airflow 3 DAGs, storing measurements and metadata in an append-only Parquet catalog. Developers can query this catalog using DuckDB SQL to generate version-pinned training manifests without re-opening raw recordings.
- • HFlow is an open-source SDK designed to process multimodal robotics data, released under the Apache-2.0 license.
- • The SDK converts raw recordings (video, joint states, actions) into standardized episodes and queryable dataset manifests.
- • It utilizes the MCAP container format to maintain synchronization across sensor streams.
- • HFlow pipelines are built using Python functions and can be executed as Airflow 3 DAGs.
- • The system stores measurements, metadata, and quality evidence in an append-only Parquet catalog queryable via DuckDB SQL.
Developers building robotics or multimodal AI applications can use HFlow to build structured, version-pinned training datasets using DuckDB SQL.
15. Keenable AI Open-Sources NEEDLE Live Search Benchmark
Keenable AI has open-sourced NEEDLE, a live benchmark for web search APIs designed to prevent overfitting by regenerating its query sets hourly or daily. Released under the MIT license, NEEDLE is a Python CLI tool that evaluates search APIs across five verticals: News, Everyday, Expert, Deep-tail, and Legal. It uses an 'ultimate' metric based on a pooled oracle engine to establish an empirical ceiling. In latency benchmarks for the week ending August 28, 2026, Keenable-realtime recorded a p50 latency of 193 ms, compared to 1,876 ms for Exa and 2,767 ms for Bing.
- • Keenable AI has open-sourced NEEDLE, a live benchmark for web search APIs that regenerates query sets hourly or daily.
- • The benchmark evaluates search APIs across five verticals: News, Everyday, Expert, Deep-tail, and Legal.
- • NEEDLE is a Python CLI tool released under an MIT license, requiring an OpenRouter key for judging.
- • Latency benchmarks for the 7-day window ending August 28, 2026, showed Keenable-realtime at 193 ms (p50), Exa at 1,876 ms (p50), and Bing at 2,767 ms (p50).
- • Deep-tail queries (representing real agent traffic) proved the most difficult, with Exa achieving 0.557 of the ultimate ceiling.
Developers building search-enabled agents can use NEEDLE to evaluate and compare the latency and retrieval quality of search APIs like Exa, Bing, and Keenable.
16. llama.cpp Changes Default Lazy-Mode Behavior for Qwen 3.8 Flash Next
A recent update to llama.cpp (b10726) alters how the runtime handles the Qwen 3.8 Flash Next model by defaulting the lazy-mode setting to auto. This configuration keeps the model's 51B-parameter PLE n-gram embedding table on disk via mmap, reading it on demand during inference. While this reduces memory usage, benchmarks show it introduces a 50% perplexity speed penalty and a 15% token generation slowdown. Developers can bypass this behavior and force the embedding table into RAM by explicitly passing the lazy-mode off flag.
- • Update b10726 changes the default lazy-mode behavior to auto for the Qwen 3.8 Flash Next model.
- • The change keeps the model's 51B-parameter PLE n-gram embedding table on disk via mmap, even with the load-mode none setting.
- • The new default lazy-mode causes a reported 50% perplexity speed penalty and a 15% token generation speed penalty.
- • Users can force the embedding table to load into RAM by using the lazy-mode off flag.
Developers running Qwen 3.8 Flash Next locally via llama.cpp need to manually disable lazy-mode to avoid a 50% perplexity speed penalty and a 15% token generation slowdown.
17. Qwen 3.8 Flash Next llama.cpp Benchmarks Reveal VRAM and Memory Layout Trade-offs
Performance testing of the unsloth/Qwen3.8-Flash-Next-GGUF model on llama.cpp b10666 highlights critical memory and layout configurations for high-end hardware. Running on an NVIDIA RTX PRO 6000 with 96GB VRAM yielded 109.07 tokens per second decode at a 2K prompt, compared to just 8.34 tokens per second on CPU-only. Crucially, forcing the 27.2 GiB per-layer token embedding table onto CUDA VRAM caused decode speeds to plummet to 1.95 tokens per second. Additionally, utilizing a non-unified KV layout achieved 92.0 tokens per second at a concurrency of 16, outperforming unified layouts.
- • Benchmarks of unsloth/Qwen3.8-Flash-Next-GGUF on llama.cpp b10666 achieved 109.07 tok/s decode with 96GB VRAM compared to 8.34 tok/s on CPU-only.
- • Forcing the 27.2 GiB per-layer token embedding table onto CUDA VRAM severely degraded decode speed from 108.5 tok/s to 1.95 tok/s.
- • RAM-resident loading provided a 1.87x higher prefill speed compared to mmap at a 48GB tensor placement.
- • At a concurrency of 16, a non-unified KV layout reached 92.0 tok/s, outperforming the unified KV layout.
- • The performance advantage of 96GB VRAM over 24GB VRAM decreased from 2.80x at 2K context to 1.45x at 245K context.
Developers optimizing local inference can achieve up to 109 tokens per second on high-end hardware with optimal memory and layout configurations.
18. Optimized Fork of BeeLlama Improves KVarN Performance at High Context Depths
Building on the initial implementation of KVarN in BeeLlama.cpp, a new GitHub fork (valujin/beellama-kvarn) introduces optimizations to address performance degradation at high context depths. Tests using Qwen3.8 IQ4 XS models on a Windows 11 system with a 5070Ti GPU show the fork achieves up to 76% faster token generation compared to the original implementation. Developers should note that dropping the -ngl flag when using kvarn still results in performance issues on both implementations.
- • The valujin/beellama-kvarn fork optimizes the existing KVarN implementation for high context depths.
- • Performance testing shows up to 76% faster token generation compared to the original BeeLlama implementation.
- • The fork was tested on a Windows 11 system with a 5070Ti GPU, 16GB VRAM, and 48GB RAM using Qwen3.8 IQ4 XS models.
- • Dropping the -ngl flag while using kvarn on both the original and forked implementations causes performance issues.
Developers using KVarN for KV-cache quantization can now mitigate performance drops during long-context inference by switching to this optimized fork.
19. llama.cpp Pull Request Speeds Up AVX2 Prompt Processing for IQ Models
A new pull request in the llama.cpp repository (#27402) introduces AVX2 optimizations designed to speed up large batch size prompt processing for IQ quantized models. This update directly improves CPU-based local inference performance for developers utilizing IQ models.
- • A pull request (#27402) by bartowski1182 in llama.cpp introduces AVX2 optimizations.
- • The update speeds up large batch size prompt processing of IQ models on CPUs.
Developers running local inference on CPUs will see improved prompt processing speeds when using IQ quantized models.