A focused feed of models, agents, research and open-source releases for people building with AI.
Partially decensored GGUF of Swift 1.5 Qwen3.8-27B built with Heretic , with vision and MTP intact. Scores 97.6% GSM8K, 86.67% MMLU-Pro, 92.98% IFEval prompt-strict on the quantized file. Follows through on roughly 1 in 5 prompts the original Swift 1.5 refused; not fully uncensored.
Artificial Analysis is exposing the CritPt v2 physics benchmark grading server through its API. CritPt covers 71 composite research challenges built by 50+ active physics researchers across 30+ institutions.
Google's Project Suncatcher ran Gemma 4 on a data-center TPU in orbit, returning a haiku to Earth. Prototype satellite named MVP, built with Planet, launched Oct 1 on SpaceX Transporter-18. Carries four Trillium-generation TPUs running on roughly a 1 kW solar power budget.
Sarvam released Prism , a 105B decision model under Apache 2.0 that outputs probabilities instead of text. Median latency of 44 ms per request on a single NVIDIA B200, 95th percentile at 420 ms. Scores 71.8 average across three decision indices, claimed highest for sub 100 ms models.
Mistral released open weights for Mistral Large 4 , a 1T-parameter MoE with 52B active params. Showcase demo: the model autonomously built Orbital Trace, a live 3D tracker of ~19,000 satellites and debris.
Prefrontal AI and ETH Zurich introduce Latent Sparse Attention (LSA), targeting the indexer bottleneck at 10M-token context At ultra-long context, the key-selection indexer consumes about 98% of a sparse attention layer's runtime LSA clusters keys hierarchically in a learned latent space so queries
Snorkel AI released RIFT , a framework that audits rubrics using nine named failure modes. RIFT identifies injected rubric defects with 75% accuracy versus 56.7% for an LLM judge baseline. Applied to 147 tasks across five public benchmarks in finance, medicine, coding, advice, and reasoning.
OpenAI unveiled LifeVal , a benchmark grading models on everyday personal-assistant tasks. Built from surveys of around 4,000 ChatGPT users plus 100 detailed task authors. Covers childcare, finances, government services, meal planning, apartment hunting, and caregiving.
Timestral 's largest model has 209 million parameters and, per Mistral, leads three pretrained zero-shot forecasting benchmarks. The smallest Timestral model forecast 96 steps from 2,048 observations in 17 milliseconds on eight CPU threads.
Parallel open-sourced PWS-Decisions , 300 classification questions whose answers depend on live web facts. Parallel Decisions API scored 76.22% vs Perplexity 50.89%, OpenAI 41.67%, Jev 33.44%. OpenAI Decisions scored 2.2x lower on PWS-Decisions than on static JevBench, exposing a live-web gap.
Qwenputer compiles a 32-bit RISC-V CPU directly into Qwen3.5-4B's weights , no external runtime. Gated DeltaNet layers become addressable RAM via one-hot keys with gate and write strength pinned to 1. Benchmarks: 100% on NUPA versus 48.6% base Qwen; 97.5% on DeepMind mul_big versus 55% base.
Mistral demoed Large 4 writing a quantum chemistry SCF solver plus 3D visualization from one short prompt. The model independently chose which bond to break and which orbitals to display. Le Chonk is a 1T parameter MoE with 49-52B active, 1M context, FP8 weights on Hugging Face.
Sakana AI will pre-train its own agent-native LLM series, Sakana Sawara , from scratch. First model, Sawara Small, targets a spring 2027 weights release; Sawara Large planned to follow. Jürgen Schmidhuber joined as Chief Scientific Advisor and keynoted Sakana's Tokyo symposium.
New arXiv paper studies how transformers internally represent sequences from multiple interacting agents. Proves any joint stochastic process factorizes into per-agent Bayesian beliefs via a feedback-loop HMM construction.
OpenGVLab released InternVL2-1B , a 0.9B parameter vision language model under MIT license. Combines InternViT-300M-448px vision encoder with Qwen2-0.5B-Instruct language backbone via MLP projector. Scores 754 on OCRBench, beating larger PaliGemma-3B (614) and Mini-InternVL-2B (654).
NVIDIA released an NVFP4 quantized DeepSeek-V4.1-Flash-1017 , roughly 400 GiB versus 475 GiB source. Accuracy on GPQA Diamond, AA-LCR, SciCode, IFBench, MMMU-Pro, and Terminal-Bench 2.1 matches source within noise.
Qwen3.5-4B-heretic drops refusals from 99/100 to 13/100 at KL divergence 0.0081 from the base model MMLU, PIQA, and WinoGrande scores stay within one standard error of the original Qwen3.5-4B Produced using Heretic 2.0 with Arbitrary-Rank Ablation on language layers 15 to 27 Ships as 2.71GB Q4_K_M
Sharp-Spark-X2.5-4B-GGUF is an imatrix-quantized repack of Spark-X2.5-4B targeting 6 GB VRAM GPUs. Q6 ties upstream Q8 on MMLU-Pro while using 43% fewer tokens at the median. Ties Anthropic Haiku 4.5 on an SWE-bench-Live slice at 4B parameters.
Amazon's chronos-bolt-mini is a 21M-parameter T5-based time series foundation model for zero-shot forecasting. Up to 250x faster and 20x more memory-efficient than the original Chronos at equal parameter count.
EdgeCIM is a CIM-based accelerator targeting the memory-bound decoding phase of decoder-only SLMs up to 4B parameters. On LLaMA3.2-1B it hits 400 tokens/s, 7.3x faster and 49.59x more energy-efficient than Jetson Orin Nano.
AA-Omniscience v1.1 shrinks to 3,000 questions across 6 domains, removing saturated and ambiguous items. Gemini 4 Carbon leads with Net Score 54, ahead of Claude Opus 5.5 (48) and Claude Fable 5.1 (46). Rank order holds with 0.98 Spearman correlation to v1, but median Net Score drops 0.6 points.
Gemini 4 Carbon (High) debuts #1 Text, #2 Agent, #4 WebDev on Arena. Carbon costs $1.80 per task, 72% cheaper than Claude Opus 5.5 at the top. GPT-6.1 Terra (Max) hits Agent Arena #8 at just $0.62 median cost per task. MAI-Image-2.7-Preview jumps to #3 in Single Image Edit, +45 pts over v2.6.
Zyphra introduces ZHA , a tree-structured sparse attention that is asymptotically subquadratic. Queries descend a tree of compressed KV summaries, skipping entire branches instead of scoring every key. Projected speedups: 35x over dense at 10M tokens, up to 417x at 1B tokens.
New preprint decomposes task vectors into Task Recognition, Task Learning, and residual components. Decomposition uses candidate-label unembeddings as the basis, keeping it linear and cheap. Bayesian optimization over two coefficients per layer beats the original task vector.
No stories match these filters.
