Latest Artificial Intelligence News & Technical Dispatches

NVIDIA ships TensorRT-LLM 0.16 with native FP4 quantization for Blackwell B200 GPUs, achieving 3.8x higher throughput and 65% memory bandwidth reduction.

Etched ships the Sohu ASIC hardwired for transformer attention, delivering 500,000 tokens per second for Llama 3 70B clusters at 8x lower serving cost.

Master SWE-bench Multimodal visual debugging benchmarks to evaluate autonomous front-end AI agents, pixel regression tests, and DOM tree layout fixes.

Master DeepSeek MLA vs standard MHA to achieve 78% KV cache compression, cutting VRAM overhead and boosting serving throughput in production LLM clusters.

NTT DOCOMO releases a dual-view adaptive Tweedie model combining retrieval-augmented RAG with compound Poisson distributions for sub-10ms edge forecasting.

Xiaomi releases MiMo v2.6 featuring 1M token multimodal context, on-device INT4 edge quantization, and sub-40ms vision reasoning for autonomous devices.

Benchmark OpenAI Triton against native CUDA C++ on NVIDIA H100 to evaluate kernel development speed, memory bandwidth saturation, and FP8 fused latency.

Benchmark BitNet b1.58 ternary 1-bit LLMs against FP16 baselines to see how replacing matrix multiplication with integer addition cuts DRAM bandwidth 71%.

Explore Cohere Command R+ Enterprise 2 featuring native multi-step tool calling, verifiable citation grounding, and private deployment on AWS Bedrock.

Anthropic launches Claude Code Enterprise featuring multi-repo codebase indexing, SOC2 isolation, and role-based permissions for autonomous coding agents.

Benchmark Eagle-2 against Medusa-2 for speculative decoding on vLLM to see how feature recycling achieves 3.4x faster token generation at zero loss.

Compare GRPO against PPO and DPO for reasoning model post-training to see how eliminating the critic network saves 42% GPU VRAM and boosts math accuracy.

OpenAI launches Operator Enterprise, delivering SOC2-compliant ephemeral browser sandboxes, automated credential isolation, and audit logging for web agents.

Mistral AI releases Mistral Large 3 featuring 256k context windows, open Apache-2.0 weights, and native multi-agent tool calling for enterprise AI systems.

Benchmark FlashInfer against FlashAttention-3 to compare FP8 tensor cores, PageAttention kernels, and sub-12ms decode latency across LLM serving engines.