1. ByteDance Launches Seedance 2.5 with 30-Second Generation and Multimodal Controls
Building on the capabilities of the Seedance line, ByteDance has officially launched Seedance 2.5. This version extends generation capacity to 30 seconds of high-quality audio-video content in a single pass and introduces advanced multimodal referencing, allowing for the simultaneous input of up to 30 images, 10 video clips, and 10 audio clips. It also adds timestamp-level editing controls and clay render referencing. API access is scheduled to be released via BytePlus ModelArk.
- • Seedance 2.5 increases generation capacity to 30 seconds of audio-video content in a single pass.
- • New multimodal referencing supports simultaneous input of up to 30 images, 10 video clips, and 10 audio clips.
- • Introduces timestamp-level editing, green screen, camera perspective editing, and clay render referencing.
- • Available on Jimeng AI and Doubao Pro, with BytePlus ModelArk API access scheduled for release.
This update significantly enhances the creative control and generation duration for developers using the Seedance platform, moving beyond the capabilities of the previous 2.0 version.
2. LongCat-Flash-Lite-Sparse Updated to Support 1M Token Context
The official weights for the LongCat-Flash-Lite-Sparse model have been released, building on the initial announcement from yesterday. This version introduces LongCat Sparse Attention (LSA) to replace the previous dense Multi-Head Latent Attention (MLA), enabling native support for 1 million tokens—a significant increase from the 256k limit reported in the initial release.
- • Model weights for LongCat-Flash-Lite-Sparse are now available for download.
- • Native context support has been expanded to 1 million tokens.
- • The architecture now utilizes LongCat Sparse Attention (LSA) instead of dense Multi-Head Latent Attention (MLA).
This update significantly increases the model's utility for processing massive documents or codebases locally by quadrupling the available context window.
3. Poolside Releases Optimized FP8 and NVFP4 Checkpoints for Laguna S 2.1
Building on the July 21 release of the Laguna S 2.1 Mixture-of-Experts model, Poolside has released updated FP8 and NVFP4 checkpoints. These new weights, accompanied by configuration modifications, provide developers with optimized options for deploying the 1-million-token context model.
- • Poolside released official FP8 and NVFP4 checkpoints for the Laguna S 2.1 model.
- • The update includes configuration modifications alongside the new weights.
- • These checkpoints are now available for download to support optimized inference.
These quantized checkpoints enable more efficient local deployment and specialized inference pipelines for the Laguna S 2.1 model.
4. Solid Queue 1.6.0 Introduces Fiber Workers for I/O-Bound LLM Workloads
Solid Queue v1.6.0 has introduced a fiber worker execution mode, allowing background jobs to run on a single fiber reactor thread instead of a traditional thread pool. This execution mode is optimized for I/O-bound workloads, making it highly suitable for applications that make frequent external LLM API calls. To use the feature, developers must configure the number of fibers, include the Async dependency, and enable fiber isolation in Rails.
- • Solid Queue v1.6.0 introduces a fiber worker execution mode that runs jobs on a single fiber reactor thread.
- • The feature is specifically intended for I/O-bound workloads, such as those involving LLM calls.
- • It requires the Async dependency and fiber isolation enabled in Rails via config.active_support.isolation_level = :fiber.
- • The release also includes a fix to roll back transactions leaked by killed job threads in tests.
This allows Rails developers to run background LLM jobs on a single fiber reactor thread instead of a heavy thread pool, improving resource efficiency.
5. llama.cpp Fixes Tool Calling Issues for DeepSeek-V4-Flash-0731
A pull request has been merged into the main llama.cpp repository to address critical tool-calling issues with the DeepSeek-V4-Flash-0731 model. The update, identified as pull request #26269, fixes model looping and poor behavior, restoring stable performance for developers utilizing the model's tool-calling capabilities in local environments.
- • Pull request #26269 was merged into the main llama.cpp repository.
- • The fix addresses issues with model looping and poor behavior during tool calling.
- • The update resolved performance issues experienced by users running DeepSeek-V4-Flash-0731.
- • The fix was added to the main branch of the repository.
This restores reliable tool-calling capabilities for developers running DeepSeek-V4-Flash-0731 locally via llama.cpp.
6. Local LLM Benchmarking Platform beta.locallm.top Launches
The beta.locallm.top platform has launched, providing a dedicated space for benchmarking local LLMs. The site currently hosts over 10 narrow-domain benchmarks and more than 25 smaller benchmarks. Developers can create custom benchmarks, define model and system prompt combinations (Pipelines), evaluate responses, and view comparison tables. The platform's architecture relies on llama-server for model execution, coordinated by a backend API, and supports private benchmarks to prevent data contamination.
- • The website beta.locallm.top was launched as a platform for benchmarking local LLMs.
- • The platform currently hosts over 10 narrow-domain benchmarks and more than 25 smaller benchmarks.
- • It supports private benchmarks and pipelines to mitigate contamination and allow for experimentation.
- • The system architecture uses llama-server for model execution, coordinated by a backend API and a Model service relay.
This gives developers a structured, private environment to evaluate local model performance on custom prompts without data contamination.
7. Cursor Removes Dollar-Based Cost Tracking from Self-Serve Dashboards
Cursor has removed dollar-based cost tracking from its usage dashboard and CSV exports for individual and self-serve Teams plans. The dashboard now displays token counts instead of dollar amounts, and historical cost data has been zeroed out or removed. Cursor stated this is an intentional design choice to prevent confusion between included plan usage and on-demand charges, though users have expressed frustration over the loss of granular per-request and model efficiency tracking.
- • Cursor removed dollar-based cost tracking for individual and self-serve Teams plans as of July 31, 2026.
- • The usage dashboard now displays token counts instead of dollar amounts, and historical cost data in exports has been removed or zeroed out.
- • Cursor representative Kevin Neilson stated the change is intended to prevent user confusion between included plan usage and on-demand charges.
- • Enterprise plans are unaffected and continue to display dollar-based cost tracking.
- • Users can check the Spending dashboard for total billing cycle costs or use the Admin API for Teams-level spending data.
This forces developers to track token counts instead of direct dollar costs, complicating budget monitoring and model efficiency analysis.