AI Briefing#benchmark
The latest AI and developer news about #benchmark, with the original source and a short summary.
The latest AI and developer news about #benchmark, with the original source and a short summary.
StepFun has released Step 3.7 Flash, a 196B-scale MoE-based multimodal model.
Open models were found to lag behind closed models by about 4 to 6 months on public benchmarks and about 8 to 10 months on private benchmarks.
AgingBench, a benchmark that measures the performance degradation of AI agents during long-term deployment, has been released.
Microsoft has released MAI-Image-2.5, an image generation model with improved text rendering and spatial awareness capabilities.
Artificial Analysis and IBM have launched ITBench-AA, a new benchmark for evaluating enterprise IT agent performance.
DeepSWE is a new coding agent benchmark with contamination-free data and high complexity.
Even the latest frontier models complete complex legal tasks perfectly less than 10% of the time.
A benchmark has been released to test whether AI can predict scientific discoveries made after its training data cutoff.
PapersWithCode has updated its platform with multi-metric support, external paper registration, and paper lineage display features.
Google Antigravity 2.0 and Gemini 3.5 Flash High achieved the highest quality scores in an OpenSCAD-based architectural 3D modeling benchmark.
A 3-billion-parameter specialized model outperformed all commercial frontier APIs, including Claude Opus and GPT-5.4, on an OCR benchmark.
Anthropic unveiled in London a coding agent that autonomously handles everything from writing PRs to deployment.
A memory system that solves the session-memory problem for AI coding agents has been released.
**VSAS-Bench**, a new framework for evaluating the performance of real-time visual streaming assistant models, has been released.
Qwen3.7-Max showed top-tier performance on coding, reasoning, and agent benchmarks.
A feature has been added to Hugging Face's benchmark dataset that allows filtering results by model size.
Cursor unveiled Composer 2.5 and gave all users 10x daily usage.
MTP speculative decoding has been added to llama.cpp, making Qwen3.6 27B up to 2.44x faster.
Ant Group released Ring-2.6-1T, a 1T reasoning model for real-world agents.
Hugging Face has released a leaderboard that compares agent systems' performance and cost together.
DystopiaBench has been expanded to 42 models and 6 dystopian scenarios to evaluate each model's ethical refusal capability.
It summarizes how to design agent evaluation harnesses that go beyond static benchmarks.
A new memory retrieval system achieved SOTA on the LongMemEval benchmark using Gemini Flash.
DeepSeek R2 has been released as open source, recording performance on par with GPT-4o.