Agent & Model Releases
Tencent Hy4 preview. Tencent released Hy4 preview, a major size increase over Hy3 with default high reasoning and a no_think mode. The community already has a ~2.38-bit official quant and a ~200GB GGUF that reportedly retains about 98% of BF16 performance.
Local Inference & Hardware
FlashMLA sm_120 port. A community build extends FlashMLA to consumer Blackwell sm_120 GPUs, showing 2–3x speedups over PyTorch SDPA in several sparse MLA workloads and 8% lower latency for FP8 KV cache decode.
Nemotron 16GB fix. Low-bit GGUFs of Nemotron-3.5-Lightning were quietly ~4.70 bpw due to row-width quantization; a patched llama.cpp build produces a real 3.07 bpw / 11.77 GiB file that runs 262K context on 16GB.
Qwen3.8-Flash-Next on Mac. A 79GB 2-bit Qwen3.8-Flash-Next ran on a 128GB M5 Max with a 350K context slot using llama.cpp, including a 333s cold prefill of a 105K prompt and ngram-mod speculative decoding.
DeepSeek V4 Flash dual GPUs. On 2x DGX Spark/GX10 setups, DeepSeek V4 Flash 0731 reaches 67–84 tok/s sustained, with local HumanEval Pass@1 of 94.5% versus 97.0% for GLM-5.3-Flash NVFP4. GLM finished the benchmark in roughly half the time.
1M context on dual 5090s. A NInfer fork adds tensor parallelism and YaRN ×4 scaling to run Qwen3.8-27B NVFP4 with 1,048,576-token context on two 5090s, decoding at 48–100 tok/s at 1M and maintaining MTP acceptance where vLLM drops to zero.
27B on 16GB. Hybrid quantization plus kvarn cache settings let Qwen3.8-27B run with 100K context at 50 tok/s on a 16GB RTX 4070 Ti SUPER, using MTP speculative decoding and tail-precision KV cache.
FreeToken MoE engine. UC Berkeley and MIT researchers introduced FreeToken, an open-source engine that dynamically co-schedules MoE expert transfers so frontier sparse models can decode on consumer GPUs without stalling on PCIe/RAM misses.
Benchmarks & Evaluations
Terminal Bench 4.0. Terminal Bench 4.0 released with a refreshed leaderboard and rapid iteration focus to combat saturation; GLM-5.3 scores at Fable 5 level within margin of error.
Finance benchmark disclosure. A benchmark card for Ling-3.0-flash-Fin shows how much agent scaffolding, tools, and evaluation pipelines shape reported results; many runs used ReAct with web search and Python, and some eval sets are internal or not yet public.
Research & Agent Infrastructure
Google WikiSkill memory. Google Research's WikiSkill gives agents a persistent wiki-like knowledge base of past failures and successes, packaging reusable Agent Skills that guide behavior without changing model weights.
Anthropic Model Hardware Standard. Anthropic is building MHS, a unified interface to let agents control physical devices like microscopes and robotic arms; early tests cut multi-machine integration from weeks to hours, but human oversight remains necessary.
Data layer for agents. A TOTVS presentation argues transactional systems and data lakes are not optimized for token-hungry, latency-sensitive agent reasoning, and describes evolving data access toward MCP and semantic models.
LAION Big Video Dataset. LAION released BVD, a research-only dataset of 10 million hours of video with 55 million clips and 300 million still images, tying video, audio and text; models trained on it outperformed InternVid baselines by up to 2.1 points on some benchmarks.
Industry & Business
Sony and Warner sue Anthropic. Sony Music Publishing, Warner Chappell and other publishers sued Anthropic, alleging torrenting and scraping of copyrighted works to train Claude and seeking damages up to $150,000 per work. Anthropic says it will defend itself.
OpenAI cuts off Cursor. OpenAI is ending its contract with Cursor after SpaceX's acquisition, citing Elon Musk companies' history of breaking contracts. Anthropic responded by positioning itself as the more reliable partner and pledged continued compute for Cursor's Claude use.
Nvidia beyond GPUs. Nvidia's earnings narrative is shifting from GPU share toward datacenter orchestration and systems around the GPU, where it has built state-of-the-art networking and orchestration hardware as AI compute scales to gigawatts.
That's everything for today - about a 5-minute read.