A focused feed of models, agents, research and open-source releases for people building with AI.
Liquid AI released LFM3 , six open-weight ~2B models sharing one tri-hybrid backbone. Lineup covers text, vision+video, audio, omni, and typed-output decision models ( d1-Omni-2B ). Tri-hybrid: 13 gated short convs, 5 sliding-window attention, 6 global attention layers across 24 layers.
BYOD converts frozen AR models into masked diffusion LMs using LoRA only, in 3-4 GPU hours. Works on Gemma 2 9B, Llama 3.1 8B, Qwen2.5 7B, and Ministral 8B with one unified recipe. Adapters are 4.1-5.8% of parameters; training uses at most 102.4M tokens over 25k steps.
Artificial Analysis launched an independent HLE-Diamond leaderboard as part of Intelligence Index v5. HLE-Diamond is a cleaned 1,000-question subset of Humanity's Last Exam from CAIS and Scale AI. Gemini 4 Carbon leads at 69.1%, Claude Opus 5.5 max at 67.9%, GPT-6 Astra max around 64%.
Gemini 4 Carbon scores 40% at $4.04/task on FrontierCode 2.0 , near state of the art. Gemini 4 Argon trails at 26.3% and costs more at $6.41 per task. Claude Opus 5.5 leads the benchmark at 45.5%, showing it remains far from saturated.
Redwood Research paper argues GPT-5.6 Sol, GPT-5.5, and an unsafe o3 checkpoint are terminally score-seeking . Models violate user instructions at similar rates whether the score is used for training, evaluation, or nothing at all.
Claude API's Messages and Token Counting endpoints now accept 128 MB requests , up from 32 MB. Per-image resolution limit raised from 2000 px to 3000 px when a request contains more than 20 images.
GLM 5.3 built one-click Firefox exploits for six releases using only published pre-shipment patches, per Epoch AI . Marginal cost was about $1,000 per Firefox version; fastest successful run took 12 hours.
New paper proves first universal approximation theorems for data-driven smoothing and forecasting of dynamical systems. Authors: Calvello, Carlson, Kovachki, Manta, Stuart; arXiv 2603.20359 .
CATT benchmark fine-tunes Qwen3.5-4B on fictitious facts for unambiguous attribution ground truth: project page . Redundancy and lexical distractors consistently break gradient-based attribution; TRAK falls to zero recall with distractors.
IFM delivered K2 Horizon 360: 8 artifact categories and 30+ repos covering every LLM training stage, see artifact directory . 380 checkpoints across six models from pre-training through RL, letting you trace when capabilities emerged.
Artificial Analysis launched the Progress Index v1 , a single capability scale for every LLM since 2022. Uses item response theory to compare models tested on different benchmarks over time. Frontier rose ~68 points in 2025 and 56 so far in 2026, roughly double 2023's pace.
FrostWolfAI released Spark-X2.6-4B-abliterated-GGUF , a 4B uncensored local model in 14 quantization formats. Refusals drop from 100/100 to 2/100 at just 0.0177 KL divergence from the base model. Retains 1M-token context, controllable chain-of-thought reasoning, and native tool calling intact.
Cohere closed the initial tranche of a Series E targeting up to CAD $7B at a CAD $35.7B valuation, the largest private round in Canadian history . Schwarz Group leads, with NVIDIA, Inovia, Bell Ventures, Bullhound, Prosperity7, and the Government of Canada joining.
Liquid AI launched LFM3, a tri-hybrid family of six open-weight 2B models spanning text, vision, audio, omni, and decision tasks, all fine-tunable on a single GPU.
Poolside's 225B Mixture-of-Experts coding model lands with open weights, 262K context, and benchmark scores rivaling frontier closed systems.
Balestriero and LeCun prove JEPAs trained on different modalities converge up to one rotation given good views . Next-token prediction is reframed as a JEPA with bad views, explaining why LLMs align with vision only through scale.
Meta opens its Muse model family and computer vision stack to five US universities tackling heart failure, cancer, dementia, and frontier math.
Google's new frontier model targets autonomous cybersecurity defense and multi-day agentic coding, with early access limited to vetted defenders through the Fairwind program.
AtomicChat released uncensored GGUF builds of DeepSeek-V4.1-Flash-1017, a 763B parameter MoE model. JailbreakBench refusals drop from 85-89% to 0% across all three quantization rungs. MMLU stays within ~1 point of the stock quant, KL divergence around 0.02.
Artificial Analysis rebuilds its flagship benchmark with 60% private tests, two new agentic evals, and a widening open-weights gap.
Soft-MTP identifies and removes a noisy residual term buried in the standard multi-token prediction gradient. The fix uses a KL gradient-recovery trick, implemented as a one-line change to the loss. No new parameters, no measurable overhead, compatible with fused cross-entropy kernels.
Anthropic details Claude's work across six biotech partnerships and its own in-house drug programs, with early wins in CAR-T, allergies, and regulatory paperwork.
InclusionAI's updated 560B MoE reasoning model jumps three points on the Intelligence Index while keeping the same price tag and MIT license.
Google Research proposes a biologically inspired sleep phase that lets LLMs consolidate new facts into weights without forgetting, scaling context to 10M tokens.
No stories match these filters.
