Brief

A new open-source tool called Reasonix has been developed to significantly reduce the cost of using DeepSeek V4 models, achieving a cache hit rate of up to 99.82%. This optimization can lower the cost of processing 400 million tokens from $61 to just $12. Reasonix is specifically designed for DeepSeek's caching mechanisms, employing an append-only loop and a cache-first strategy to keep older contexts stable and minimize recomputation. The tool also includes features for repairing tool calls and intelligently managing model versions to further control expenses. AI

DeepSeek-V4 introduces novel training techniques, including Anticipatory Routing to stabilize training by using older weights for routing decisions, and a Generative Reward Model (GRM) where the model itself acts as a judge for complex tasks. The model also supports three distinct reasoning modes (Non-think, Think High, Think Max) trained with varied configurations for different reasoning depths. These advancements highlight the need for flexible, programmable training infrastructure that can adapt to complex, co-designed model and runtime systems. AI

A new open-source coding-focused language model, MiMo-V2.5-coder, has been released. The model is presented as a strong alternative to Qwen3.6 and DeepSeek-V4, particularly for coding tasks. It is noted for its speed and reliable tool-calling capabilities, requiring 128GB of RAM. AI

A new benchmark called SaaS-Bench has revealed that current AI agents struggle significantly with real-world, long-horizon tasks, with top models like Claude Opus 4.7 achieving less than 4% success rate on fully completing tasks. The benchmark uses actual SaaS systems and data, exposing four key failure modes: inability to maintain performance over extended tasks, cascading errors from single mistakes, a lack of self-checking mechanisms, and inconsistent performance across multiple runs. These findings suggest that the current paradigm for AI agents is insufficient for true automation and that software interfaces may need to be redesigned for AI agents rather than expecting them to operate human-centric interfaces. AI

A comparison of Claude Opus 4.5 and DeepSeek V4 highlights their distinct strengths in coding tasks. Claude Opus 4.5 excels at precise, surgical fixes for production bugs and single-file issues, achieving a leading 80.9% score on the SWE-bench benchmark. DeepSeek V4, conversely, is better suited for large-scale, multi-file refactoring and repository-wide migrations when provided with extensive context. The choice between them depends on the scope and nature of the coding task. AI

DeepSeek V4 Flash, a new iteration of the DeepSeek V4 model, has demonstrated impressive performance metrics. It achieves a throughput of 350 tokens per second with a latency of approximately 1.5 seconds. This advancement is attributed to Tensorix and Cortecs, with implications for AI development in the EU. AI

The latest release of the Hermes Agent Foundation provides access to the DeepSeek V4 model and introduces native Windows support. This update aims to improve accessibility and usability for users. The release details were shared by World of AI. AI

Opencode Go offers a coding environment using open-source models like DeepSeek V4. The service costs €5 for the first month, then €10 per month, with a €5 discount available. AI

A user shared their experience integrating Deepseek V4 models into the Cursor IDE via an OpenAI-compatible API. Initial attempts with direct Deepseek API keys resulted in errors on longer prompts, while using Openrouter's API led to slow performance and high token consumption. The user found better results using the Cline plugin with an Openrouter API key, but concluded their direct API integration experiments were largely unsuccessful. AI

Users of the AI-powered code editor Cursor are expressing concern over potential changes to its pricing model. Some users are worried that Cursor might adopt a usage-based pricing system, similar to what they've observed with other AI tools like Codex and Claude. This shift would move away from their current flat monthly subscription, which is seen as more predictable and cost-effective for heavy users. AI

Hugging Face has launched new features for PapersWithCode, a platform tracking AI state-of-the-art. The updates include support for multiple metrics on leaderboards, such as for Automatic Speech Recognition and Object Detection. The platform now also accommodates external papers beyond arXiv, automatically enriching them with relevant tags and data, and displays paper lineage to show follow-ups or predecessors. AI

Together AI has detailed the architectural innovations behind DeepSeek-V4's ability to handle a 1 million token context window. The model employs a hybrid attention design that compresses context before storing it in the KV cache, significantly reducing memory pressure. This architectural shift transforms the challenge of long-context inference from a model capability into an inference systems problem, requiring optimized serving engines to manage cache layouts and batching effectively. AI