OpenBMB's MiniCPM5-2B Tops Sub-4B Charts While Burning 3x Fewer Tokens
OpenBMB's 2.6B dense reasoning model tops the sub-4B open weights leaderboard, punching well above its weight on agentic benchmarks while staying token-efficient.
// Intelligence
Llms
A focused feed of models, agents, research and open-source releases for people building with AI.
Updated
Stanford and CMU Link Character.AI Use to Lower Well-Being Over Time
A year-long Stanford study of 1,182 Character.AI users finds heavier chatbot engagement predicts lower well-being, largely by crowding out face-to-face contact.
TokenRhythm's NeoHorse-1 Beats Its Base Model by 6 Points on Agentic Benchmarks
TokenRhythm fine-tunes Qwen3.5-4B with a routing harness that turns agent traces into training data, gaining 5.93 points across ten benchmarks.
Japan's LLM-jp-4-VL 9B Beats InternVL on Japanese While Staying Truly Open Source
Japan's national LLM project releases an Apache-2.0 vision-language model with reasoning traces, trained on a scrubbed 29M-sample dataset.
OpenBMB's MiniCPM5-2B Beats Larger Rivals With 128K Context on Consumer Hardware
OpenBMB's compact 2B model hits open-source SOTA in its class, beating several 4B rivals on code, math, and agent benchmarks while running locally.
Artificial Analysis Rebuilds Its AI Leaderboard to Stop Models From Gaming It
Artificial Analysis pushes Intelligence Index v4.2 with private test sets, a 4,592-page PDF reasoning benchmark, and a new agentic knowledge-work suite.
MBZUAI Releases K2 Horizon, the Largest Fully Open-Source AI Fleet Ever
IFM released six fully open models from 0.9B to 375B parameters, complete with weights, code, training data, and recipes under Apache 2.0.
Ant Group's Ling-3.0-flash-VL Brings Images and Video to Its 124B Open Model
InclusionAI open-sources a 124B-parameter mixture-of-experts vision-language model with 5.5B active parameters, 256K context, and MIT license.
OpenAI's GPT-6 Astra Hits 99.9% on AGI Benchmarks and Runs Your Computer
OpenAI's new flagship saturates ARC-AGI-3 and FrontierMath Tier 4, drives a browser at 1.9x speed, and crosses the Critical cybersecurity threshold.
Coding Agent Cost: Why Sessions Get Expensive
Keep the project on disk, treat lunch as a bill, and score the first finished tasks rather than dollars per million tokens
Google DeepMind's WeatherNext 3 Ditches Physics Simulations for 60% Sharper Forecasts
Google DeepMind's new weather AI trains on raw satellite feeds and station data, delivering hourly 5km forecasts with up to 50% better rain accuracy.
Xiaohongshu's Self-GC Cuts Agent Token Costs by 20% Without Losing Critical Context
Xiaohongshu's Self-GC uses a planner LLM to fold, mask, or prune agent context, retaining critical details 84.85% of the time versus 54.55% for rule-based baselines.
Researchers Crack Open GPT-OSS-20b to Find Hidden Symbol Systems Inside
An 8-year interpretability project shows that LLM representations can be closely approximated by symbolic role-filler structures, enabling precise behavioral edits.
Google DeepMind's Gemini 3.8 Flash Beats Claude Opus 5 at 6x Lower Cost
Google's third Flash release in as many months pushes a workhorse model into frontier-tier territory on agentic coding, legal, and finance benchmarks.
Multiverse Computing's Quasar 438B Tops Europe's AI Leaderboard While Beating Bigger Models
Multiverse Computing's Quasar 438B tops European AI rankings with a 43 Intelligence Index score and 15.3-second reasoning responses.
Alibaba's Qwen3.8-Max-0902 Tops Code Arena, Beating Claude Opus 5
Alibaba's flagship 2.4T-parameter model gets a coding and agent-work refresh with the same pricing and 1M context window.
Anthropic's Claude Fable 5.1 Tops Every Frontier Model but Costs 20% More
Independent evaluations put Anthropic's new flagship at the top of the intelligence charts, but the win comes with a token-count tax that eats the cache savings.
Microsoft Finds a 2023 Trick Beats Two Years of Linear Attention Research
A new benchmark study finds that sliding-window attention with sinks matches or beats post-trained linear attention models, without any retraining.
Cursor's Claude Fable 5.1 Hits 73.4% and Catches Its Own Bugs
Cursor added Anthropic's new Fable 5.1 model, which posted a 73.4% score on CursorBench 3.2 and excels at self-verifying long coding runs.
Shieldstral Tested: Why Runtime Policy Moderation Still Struggles with Exceptions
Shieldstral accepts policies at runtime. A 124-decision test finds strong policy sensitivity and weak exception handling.
Tsinghua's SMELT Cuts AI Training Costs 18% by Looping Layers Twice
New scaling laws show that looping the middle half of a Mixture-of-Experts model twice saves up to 18% of training compute at matched budgets.
DavidAU's Qwen3.8 Fine-Tune Beats the Base Model at a Tenth of Thinking Tokens
A community fine-tune of Qwen3.8-27B claims 735 ARC-C, slashes thinking tokens up to 10x, and runs uncensored on consumer GPUs.
Google's TimesFM-3 Tops Three Forecasting Leaderboards With 330M Parameters
Google's new 330M-parameter time series foundation model handles multivariate forecasting zero-shot in a single forward pass, topping Gift-Eval, FEV-Bench and Time.
What a Week of AI Security Incidents Means for Developers
Seven practical shifts for securing agents that can find any crack
No stories match these filters.