1. Anthropic Releases Claude Opus 5 with Adjustable Thinking and Relaxed Code Safety
Anthropic has released Claude Opus 5 as its new flagship model, replacing Claude Opus 4.8. The model is designed for bounded, complex tasks and serves as the new default for the Claude Max consumer tier. It approaches the capabilities of Claude Fable 5 in many domains while performing better at complex coding tasks. To address safety, Anthropic has relaxed safeguards to allow vulnerability finding in source code, while maintaining restrictions on binary-based scanning and exploit generation. The model also features lower rates of deceptive behavior and decreased indirect prompt injection success rates. Additionally, Anthropic is working with government partners to conduct independent testing of the model.
- • Claude Opus 5 is priced at $5 per million input tokens and $25 per million output tokens, matching its predecessor Opus 4.8.
- • The model features a 1M-token context window and supports up to 128k output tokens on the synchronous Messages API.
- • Thinking is enabled by default, with an adjustable effort parameter that controls reasoning depth; setting max_tokens with effort above 4096 returns a 400 error.
- • Anthropic has relaxed safety safeguards to allow source-code vulnerability scanning, while maintaining restrictions on exploit generation.
- • New API features include a Fast mode for double the price, automatic fallback routing, and support for mid-conversation tool changes.
Developers get a highly capable, cost-efficient flagship model with native reasoning controls and fewer restrictions on source-code security analysis.
2. Microsoft Extends MAI-Image-2.5-Pro and MAI-Voice-2-Flash to Public Preview
Building on the recent integration of MAI-Image-2.5-Pro and MAI-Voice-2-Flash into core products like PowerPoint and Bing, Microsoft has now made these models available to developers in public preview. This release expands access to the high-fidelity image generation and low-latency voice capabilities previously reserved for internal product use.
- • MAI-Image-2.5-Pro and MAI-Voice-2-Flash are now available in public preview for developers.
- • This follows their recent deployment into Microsoft consumer products like Bing and PowerPoint.
- • The models are now part of Microsoft's production model lineup for external use.
Developers can now integrate the same high-performance models powering Microsoft's core consumer applications into their own software.
3. CachyLLama Fork Optimizes Local Agent Sessions with Persistent KV Caching
CachyLLama is a fork of llama.cpp designed to optimize repeated prompt processing on slower hardware. The project features persistent SSD-backed KV checkpoints and a system-prompt cache to restore previously processed context. CachyLLama checkpoints are capable of surviving server restarts. The software does not increase generation speed but reduces prompt-evaluation time by avoiding redundant work.
- • CachyLLama is a llama.cpp fork featuring persistent SSD-backed KV checkpoints and a system-prompt cache that survives server restarts.
- • The software reduces prompt-evaluation times, achieving warm processing times of 0.99s for 15,700 tokens on a 7840U system.
- • It supports hybrid architectures including Qwen 3.5/3.6, Gemma 4, and GLM-4.7.
Developers running local agents can eliminate redundant prompt-processing overhead and maintain context across server restarts.
4. Build High-Resolution OCR Pipelines with Baidu's 3B Unlimited-OCR Model
A new tutorial demonstrates how to build an end-to-end OCR pipeline using Baidu's 3B-parameter Unlimited-OCR vision-language model. The workflow is designed to run in Google Colab on CUDA-enabled GPUs, with automatic selection of bfloat16 or float16 precision based on hardware support. The pipeline includes settings for long-context generation, repetition controls, and structured output handling to produce text, Markdown, MMD, and JSON artifacts.
- • The pipeline uses Baidu's 3B-parameter Unlimited-OCR vision-language model running on CUDA-enabled GPUs with automatic bfloat16/float16 selection.
- • It supports two inference modes: Gundam mode for tiled image crops in dense layouts, and Base mode for a single 1024-pixel view.
- • Multi-page PDFs are parsed by rasterizing pages into PNGs using PyMuPDF and processing them with the infer_multi() function.
- • The pipeline outputs structured text, Markdown, MMD, and JSON artifacts with controls for long-context generation and repetition.
Developers can implement a high-resolution, multi-page OCR pipeline locally or in Colab with structured Markdown and JSON outputs using a lightweight 3B model.
5. SupraLabs Releases 5-Million-Sample Reasoning Corpus for Fine-Tuning Small Models
SupraLabs has released a reasoning corpus dataset containing 5 million samples on Hugging Face. The dataset is designed to facilitate supervised fine-tuning (SFT) and training of tiny small language models (SLMs). All samples are limited to a 5,000-token sequence length to ensure compatibility with low-resource training setups.
- • The dataset contains 5 million samples and is hosted on Hugging Face as SupraLabs/reasoning-corpus-4K-5M-v1.
- • Each sample includes a user prompt, model thought trace, assistant answer, repository ID, and total token length in ChatML format.
- • All samples are capped at a 5,000-token sequence length to facilitate supervised fine-tuning (SFT) of tiny models.
Developers can use this structured reasoning corpus to fine-tune small, task-specific models with explicit thought traces and a 5,000-token sequence limit.
6. Noema Overfit Enables Gemma 4 26B to Run on iPhone via SSD Model Paging
Noema has released a feature called Noema Overfit in the Noema app. Noema Overfit enables large models, such as the Gemma 4 26B A4B (Q4_K_M version), to run on devices like the iPhone 17 Pro using a paging technique. The paging technique stores non-expert weights in RAM while reading model experts from the SSD. The system is designed for scenarios where answer accuracy is prioritized over speed and is also suitable for low-RAM MacBooks.
- • The Noema Overfit feature in the Noema app enables the Gemma 4 26B A4B model to run on devices like the iPhone 17 Pro.
- • The paging technique stores non-expert weights in RAM while dynamically reading model experts from the SSD.
- • Testing on an iPhone 17 Pro yielded a prefill speed of 34.4 tokens per second and a decode speed of 3.5 tokens per second.
- • The system is designed for high-accuracy scenarios where speed is secondary, and is also compatible with low-RAM MacBooks.
Developers can deploy larger, more capable open-weights models directly to mobile and low-spec consumer hardware by leveraging SSD-backed weight paging.
7. Hetzner Launches Experimental Free LLM Inference API for Qwen MoE Model
Hetzner has launched an experimental LLM inference service that is currently free and lacks SLAs or production guarantees. The service provides an OpenAI-compatible API for the Qwen/Qwen3.6-35B-A3B-FP8 model, which is a 35-billion-parameter Mixture-of-Experts model with 3 billion active parameters, a 262K context window, and FP8-quantized weights. Hetzner's public dedicated GPU server offerings currently consist of NVIDIA RTX 4000 SFF Ada Generation and NVIDIA RTX PRO 6000 Blackwell Max-Q GPUs.
- • Hetzner's experimental LLM inference service is currently free and provides an OpenAI-compatible API without SLAs.
- • The service hosts the Qwen/Qwen3.6-35B-A3B-FP8 model, a 35B MoE model with 3B active parameters and a 262K context window.
- • Benchmark tests recorded a median time to first token of 153 ms and an output speed of 224 tokens per second.
- • The API includes an undocumented enable_thinking parameter to toggle the model's reasoning budget.
- • Hetzner's public dedicated GPU server options currently feature NVIDIA RTX 4000 SFF Ada and RTX PRO 6000 Blackwell Max-Q GPUs.
Developers can test a highly responsive, quantized 35B MoE model for free with an OpenAI-compatible endpoint before deploying on dedicated GPU hardware.