Moonshot AI's Kimi K2.6 model card is live on the Hub, a 1T-parameter MoE model with a DeepSeek V3-style architecture and native text, image, and video input, released under a modified MIT license.

New paper proposes an SVD-based spectral-energy-guided bit allocation scheme for quantizing Mixture-of-Experts LLMs, claiming a 12.3x faster quantization pass and 1.76x decoding speedup over GPTQ at 2-bit precision on Qwen3-30B-A3B.

OpenAI shared plans for Jalapeño, a custom inference chip built with Broadcom, joining Google, Apple, and SpaceX in reducing reliance on Nvidia for AI compute.

Google's Gemini Embedding 2 is now generally available, mapping text, images, video, audio, and PDFs into one embedding space with Matryoshka Representation Learning for flexible 3072/1536/768-dim output.

The contamination-free LiveCodeBench v6 leaderboard now covers 53 evaluated models across code generation, self-repair, execution, and test-output prediction, with Qwen3.7 Max leading at a 0.916 score.

Z.ai's open-weight GLM-5.2 outperforms GPT-5.5 on several long-horizon coding benchmarks at roughly 1/6th the cost, a good reminder to weigh cost-per-task alongside raw accuracy when designing eval harnesses.

Anthropic's engineering team argues the discipline has moved past prompt wording alone: the real lever is curating the smallest set of high-signal tokens across the whole context window, not just the system prompt.

A paper proposing lookahead retrieval to prefetch likely-needed documents before the generation step even asks for them, trimming end-to-end RAG latency without hurting answer quality.

A practical walkthrough of fine-tuning a small open model end to end in PyTorch and Hugging Face Transformers, from tokenization through training loop to eval.

Databricks breaks down how rank, alpha, and target-module choice actually trade off against VRAM and overfitting risk when you're doing LoRA fine-tunes.