1. Tencent Releases Hy4 Preview with 1M Context Window
The Hy4 preview is designed for productivity tasks including software engineering, game development, and scientific research. In internal evaluations on engineering tasks, the model scored 2.99 out of 4.00, outperforming GLM-5.3 and Kimi K3. Additionally, the model participated in its own development by autonomously optimizing its inference system, which yielded a 31.8% increase in end-to-end throughput.
- • The Hy4 preview features 770 billion total parameters, 49 billion active parameters, and a context window exceeding 1 million tokens.
- • API pricing is set at $0.834 per million input tokens, $2.501 per million output tokens, and $0.042 per million tokens for cache hits.
- • The model is available globally via Tencent Cloud TokenHub, OpenRouter, WorkBuddy, and CodeBuddy.
- • Tencent compressed the model from 1.5TB to approximately 200GB in GGUF format while retaining about 98% of its original performance.
Developers can access a highly capable, long-context mixture-of-experts model globally via OpenRouter and Tencent Cloud at highly competitive API pricing.
2. vLLM v0.28.0 Adds DeepSeek V4 Support and Doubles Batched Tokens
This major release incorporates 584 commits from 270 contributors. In addition to model and hardware support expansions, the Transformers library dependency has been updated to version 5.15.0, and bitsandbytes support has been migrated to an out-of-tree plugin. Pre-built wheels are available for CUDA 12.9 and 13.0, as well as CPU-only environments including macOS.
- • vLLM v0.28.0 introduces DeepSeek V4 support, featuring sparse MLA, AMD Quark NVFP4 support, and ROCm enablement.
- • The default max_num_batched_tokens has been doubled from 8,192 to 16,384 to improve throughput.
- • Performance optimizations for Kimi-K3 include Decode Context Parallel support, fused FlashKDA kernels, and an adaptive speculative token budget.
- • The release adds support for new models including Muse Glimmer, Ling 3.0 Flash, Dots3 NOTE, and Interns2mobius.
Developers hosting open-weights models can leverage optimized DeepSeek V4 serving and double their default batched token capacity for higher throughput.
3. TerminalBench 4.0 Launches to Reduce Agent Evaluation Costs
Building on previous versions of the Terminal-Bench suite, TerminalBench 4.0 introduces a lightweight framework designed to address the high costs and token requirements of earlier iterations. It allows developers to iterate quickly on agent harnesses and tool integrations without the need for the massive token counts required by previous benchmarks.
- • TerminalBench 4.0 succeeds previous versions of the Terminal-Bench series.
- • The new version reduces the token requirements for benchmarking, moving away from the 5-10 billion token scale.
- • It is optimized for rapid iteration of custom coding agent harnesses.
- • The release provides a more accessible way to evaluate tool modifications and prompt engineering.
This update enables developers to measure the impact of agent modifications on performance and token efficiency more affordably, facilitating faster development cycles.
4. StemDeck Releases Open-Source Local AI Audio Stem Separator
StemDeck provides a complete local environment for audio manipulation, featuring a DAW-style multitrack mixer, a waveform editor, and integrated analysis tools for BPM, key, scale, and loudness (LUFS). By operating entirely on-device, it eliminates the need for subscriptions, accounts, or cloud uploads, making it a secure and cost-effective option for local audio preprocessing.
- • StemDeck separates audio files or YouTube URLs into six tracks: vocals, drums, bass, guitar, piano, and other.
- • The tool is built on Python 3.12 with a FastAPI backend and a Tauri v2 desktop shell, running entirely locally.
- • It utilizes the Demucs htdemucs_6s neural network for high-quality audio separation.
- • The application supports GPU acceleration via CUDA on NVIDIA hardware and MPS on Apple Silicon across macOS, Windows, and Linux.
Developers building audio-processing or multimodal applications can leverage a fully local, GPU-accelerated pipeline to isolate vocals and instruments without cloud dependencies.
5. Claude Code Sets Permanent Weekly Usage Limits Following Temporary Boost
Building on previous limit expansions for Claude Code, ClaudeDevs has announced a new permanent policy. Starting September 14, users on Pro, Max, Team, and Enterprise plans will see a permanent 25% increase in standard weekly limits. This replaces the temporary 50% boost currently in effect, representing a 17% reduction from the temporary levels but establishing a higher long-term baseline for developer workflows.
- • Permanent 25% increase to weekly limits takes effect September 14 for Pro, Max, Team, and Enterprise plans.
- • The change transitions users from the current temporary 50% limit increase.
- • The new permanent limit represents a 17% reduction from the temporary 50% levels.
- • This follows previous limit increases for Claude Code reported in May and June 2026.
This update formalizes the long-term capacity for Claude Code users, providing a permanent increase over the original baseline established in earlier limit expansions.