AI developer tools · What shipped

Mistral launched a public preview of Mistral Large 4 (ML4), a 1 trillion-parameter natively multimodal model with 49 billion active parameters. The model demonstrates competitive performance with leading open-source models and outperforms all open-weight models from the US or Europe. ML4 achieves state-of-the-art results on multiple benchmarks including the Artificial Analysis Cyber Index (ranking top five globally for cybersecurity), Coding Agent Index (49.8%), AutomationBench (59.9%), and SciCode-Verified for scientific tasks. The model shows particular strength in visual grounding, surpassing GPT-6-Astra on Dense 200 (42% vs 41%), and outperforms competitors on legal and financial benchmarks. Model weights will be released by end of month, with the preview API available now on Mistral Studio. The model was trained on 3,800 NVIDIA Grace Blackwell GPUs in Mistral's European datacenters and incorporates reinforcement learning at scale, generating approximately 33 billion tokens per training run day with 16 billion trainable completion tokens after filtering.

Mistral released a preview of Mistral Large 4, a 1-trillion-parameter open-weight AI model, with API access available immediately and model weights launching October 27. The model processes text and images, trained from scratch over two months using approximately 4,000 Nvidia Grace Blackwell GPUs in European data centers. On the DeepSWE agentic coding benchmark, Large 4 scored 62%, ahead of competitors like Zhipu's GLM-5.3 (61%) and DeepSeek-V4-Pro (57%). The model achieved 67% on the FinWorkBench finance benchmark and demonstrated particular strength in vision tasks, scoring 73% on the DIOR-RSVG remote-sensing grounding test compared to 68% for GPT-6 Astra. Mistral emphasizes cybersecurity capabilities and sovereignty, positioning the open-weight model as advantageous for enterprises and governments that need to run AI systems on their own infrastructure without external dependency.

llama.cpp released multiple pre-release updates with compiler and backend optimizations. Recent changes include CUDA BF16 support for the XIELU kernel with testing across Mac CPU, Metal, and RTX 4090 platforms; K2 Horizon dense and MoVA model support with unicode regex fixes for proper pre-tokenization; and CLAMP operation fixes for non-contiguous views on CPU and CUDA backends using fastdiv for view strides. Builds are distributed across macOS, Linux, Android, Windows, and openEuler platforms with multiple backend options including CUDA 12/13, ROCm, Vulkan, OpenVINO, and SYCL.

Burn 0.22 introduces substantial compiler optimizations and framework improvements for machine learning development. Removing backend type parameters from the user-facing API reduces compilation dependency chains, achieving rebuild speedups up to 15× faster—specifically 6.22× faster for CNN model edits and 14.73× faster for transformer custom loops. The release includes adaptive memory management for CubeCL with peak VRAM reductions of 49.2% for CNNs and 17.7% for transformers, improved GPU training performance with 44.3% step-time reduction and 1.80× throughput increase on NVIDIA GPUs, and smarter autotuning via graph capture and replay on CUDA and HIP. Additionally, Burn 0.22 adds LoRA and QLoRA support for efficient model fine-tuning, ONNX export capabilities, and expanded remote compute with Iroh-based peer connections, representing significant updates to the machine learning framework's developer tooling and execution efficiency.