Tagged articles
benchmark performance
27 articles · Page 1 of 1
The paper systematically derives scaling laws for mixture-of-experts diffusion language models, trains a 30B-parameter LLaDA MoE v2 model guided by these laws, and demonstrates that it matches or exceeds state-of-the-art autoregressive models while using far less compute.
LLaDA MoE v2benchmark performancediffusion language model
0 likes · 10 min read

BigBang‑V1 demonstrates that a 35‑billion‑parameter model can surpass much larger models by using a two‑level generator‑critic framework that lets AI create, verify, and iteratively improve its own training tasks, achieving top scores on multiple scientific and code benchmarks without scaling model size.
BigBang-V1benchmark performancegenerator‑critic loop
0 likes · 14 min read

The article analyzes why most multi‑agent AI projects fail, introduces EvoX’s swarm‑self‑evolution approach that splits tasks into atomic units, shows benchmark results where EvoX rivals Codex while cutting per‑task cost to $1.95, and explores how information design drives agent self‑organization.
AI agentsEvoXbenchmark performance
0 likes · 12 min read

Daxiao Robot’s Kairos architecture unifies multimodal understanding, generation, and prediction in a single native design, outperforms Nvidia’s Cosmos 3.0, tops four global embodied‑AI benchmarks, and achieves real‑time edge deployment through a novel training curriculum and hardware‑aware optimizations.
Edge DeploymentEmbodied AIKairos
0 likes · 12 min read

VibeThinker‑3B, a 3‑billion‑parameter language model built on Qwen2.5‑Coder‑3B, achieves scores within the range of 671 B‑parameter models on benchmarks such as LiveCodeBench, AIME26, IMO‑AnswerBench and GPQA, thanks to a two‑stage SFT, multi‑domain reinforcement learning, offline self‑distillation and a claim‑reliability (CLR) evaluator that together push its reasoning ability to the frontier.
Post-TrainingVibeThinker-3Bbenchmark performance
0 likes · 9 min read

VibeThinker‑3B, a 3‑billion‑parameter Chinese‑built model, achieves programming benchmark scores comparable to top‑tier models like Opus 4.5, excelling in AIME, HMMT, LiveCodeBench and LeetCode contests, thanks to its Spectrum‑to‑Signal training pipeline, Claim‑Level reliability evaluation, and multi‑stage SFT and RL refinements.
AI researchClaim-Level ReliabilitySpectrum-to-Signal
0 likes · 7 min read

Gemini 3.2 Flash was quietly released on the web, instantly generating thousands of lines of code—including a 2200‑line Three.js demo and a full Windows 98 replica—thanks to model distillation and sparsification that deliver near‑GPT‑5.5 performance with 15‑20× lower inference cost, while also integrating with services like Canva, Instacart and OpenTable ahead of the I/O 2026 conference.
AI code generationGemini 3.2Google AI
0 likes · 8 min read

Qwen3.7-Plus is a multimodal large‑model that unifies vision and language, delivers top‑5 global Vision Arena rankings, excels on a wide range of pure‑text, visual‑reasoning, and video benchmarks, and powers autonomous agents that perceive screens, generate code, and complete complex GUI/CLI workflows end‑to‑end.
Agent Automationbenchmark performancecode generation
0 likes · 14 min read

At Google I/O 2026 the company launched Gemini Omni—a multimodal model that creates video from any input—alongside Gemini 3.5 Flash, which outperforms its predecessor on every benchmark, introduced the Antigravity 2.0 agent platform capable of building an OS from 93 agents, and debuted Gemini Spark, a 24/7 personal AI assistant, while also revealing pricing and upcoming releases.
AI agentsGemini 3.5 FlashGemini Omni
0 likes · 12 min read


ArcThinkMay 29, 2026 · Artificial Intelligence
Claude Opus 4.8, released on May 28 2026, keeps the same 1 M‑token hybrid reasoning model and pricing but adds modest benchmark gains, stronger honesty in code‑summary reporting, Dynamic Workflows for multi‑agent orchestration, a more complex cost structure, and new security considerations, guiding engineers on when and how to adopt it for high‑value, long‑running tasks.
AI agentsClaude Opus 4.8benchmark performance
0 likes · 17 min read

The Chinese Academy of Sciences unveils SpikingBrain 2.0‑5B, a brain‑inspired large model that uses dual‑space sparse attention and dual activation (FP8 and INT8‑Spiking) to cut training cost by over tenfold, achieve up to 15× speedup on long sequences, and match Qwen‑3 performance while drastically reducing power consumption.
SpikingBrain2.0benchmark performancebrain-inspired AI
0 likes · 10 min read

DeepSeek‑V4, released on April 24 2026, offers open‑source Pro and Flash versions with 1 M‑token context, benchmark‑leading performance, advanced agent capabilities, sparse‑attention efficiency, competitive pricing, and flexible deployment options for developers, enterprises, and content creators.
1M contextAgent CapabilitiesOpen Source LLM
0 likes · 7 min read

Anthropic’s Claude Mythos Preview uncovered thousands of zero‑day bugs across major operating systems and browsers, outperformed all benchmark suites, and is being kept out of the public market in favor of a exclusive Project Glasswing partnership with twelve tech giants.
AI securityAnthropicClaude Mythos
0 likes · 11 min read

The article details how MiniMax's M2.7 model, equipped with native multi‑agent collaboration and a 97% instruction‑following rate, autonomously executes an end‑to‑end research workflow—discovering topics, generating experiment roadmaps, fixing bugs, and achieving up to 30% performance gains and a 66.6% Kaggle medal rate—demonstrating a practical leap from benchmark scores to real‑world engineering reliability.
AI agentsKaggle MLE LiteMiniMax-M2.7
0 likes · 9 min read


AIWalkerMar 19, 2026 · Artificial Intelligence
Vision‑R1 introduces a 7B multimodal large language model that leverages 200K unsupervised CoT data, Modality Bridging, and Progressive Thinking Suppression Training to overcome data scarcity and over‑thinking, achieving 73.5% accuracy on MathVista—within 0.4% of OpenAI’s O1.
benchmark performancechain-of-thoughtlarge language models
0 likes · 12 min read

Alibaba's Tongyi team released the Qwen3‑Max‑Thinking model, surpassing one trillion parameters and 36 T tokens of pre‑training, introducing adaptive tool‑calling and test‑time expansion techniques that boost benchmark scores across 19 tests, positioning it alongside top international models and now available via app, web, and API.
AI Model ComparisonQwen3-Max-Thinkingadaptive tool calling
0 likes · 5 min read

The article examines how open‑source models like DeepSeek‑R1 and Kimi K2 Thinking are challenging the traditional closed‑source, high‑capital AI paradigm by achieving comparable or superior benchmark results at a fraction of the training cost, reshaping market expectations, investment strategies, and the economics of AI development.
AI market dynamicsbenchmark performancemixture of experts
0 likes · 11 min read

Kwai's newly released Keye-VL-1.5-8B multimodal large language model dramatically improves visual, reasoning, and temporal understanding, achieving top scores on public video benchmarks and surpassing closed‑source models like GPT‑4o, while offering an open‑source release and detailed technical documentation.
benchmark performancemultimodal LLMopen-source
0 likes · 11 min read

Fast‑search platform Kwai has open‑sourced the 8‑billion‑parameter multimodal LLM Keye‑VL‑1.5, which introduces a slow‑fast frame encoding, a progressive four‑stage pre‑training pipeline, and an automated data construction workflow, achieving state‑of‑the‑art results on video and vision‑language benchmarks and surpassing many closed‑source models.
benchmark performancelarge language modelmultimodal AI
0 likes · 12 min read

LongCat‑Flash‑Chat, Meituan’s newly open‑sourced 560B MoE model, outperforms leading LLMs on agent tool use and instruction following benchmarks, introduces zero‑computation experts and shortcut‑connected MoE for higher throughput, and demonstrates strong programming and reasoning abilities across diverse evaluation tasks.
Meituan AIZero Computation Expertsbenchmark performance
0 likes · 12 min read

Kimi K2, an open‑source trillion‑parameter AI model released by Moonshot AI, offers Base and Instruct variants, achieves leading scores on benchmarks such as SWE‑bench, LiveCodeBench and AceBench, and introduces a novel post‑training autonomous‑exploration stage with MuonClip optimization to enable robust tool use and reinforcement‑learning‑driven self‑improvement.
Kimi K2autonomous agentsbenchmark performance
0 likes · 8 min read

The article examines Kimi‑Researcher, an AI research agent built with end‑to‑end reinforcement learning, detailing its technical motivations, advantages over traditional workflow‑based and SFT methods, performance breakthroughs on benchmark exams, and diverse real‑world use cases ranging from literature reviews to legal analysis.
AI agentEnd-to-End RLKimi Researcher
0 likes · 10 min read

Flux Kontext, Black Forest Labs' new open‑weight AI image editing suite, enables fast, low‑cost contextual generation and editing with features such as role consistency, local edits, style transfer, and superior benchmark performance compared to GPT‑Image‑1, Imagen 4, and other leading models.
AI image generationFlux Kontextbenchmark performance
0 likes · 12 min read


AIWalkerApr 13, 2025 · Artificial Intelligence
Huawei's Pangu Ultra introduces a 135‑billion‑parameter dense language model trained entirely on Ascend NPUs, detailing novel stability architectures, a domain‑aware tokenizer, multi‑stage pre‑training, extensive system optimizations, and benchmark results that surpass leading models such as Llama 405B and DeepSeek‑R1.
Ascend NPUDense ModelSystem Optimization
0 likes · 15 min read


21CTOMar 27, 2025 · Artificial Intelligence
Google's Gemini 2.5, billed as its most intelligent AI model, introduces advanced reasoning capabilities that outperform rivals on benchmarks like LMArena and Humanity's Last Exam, excels at web and agent code generation, and is now available to premium users via AI Studio with a 1‑million token context window.
AI ReasoningGoogle Geminibenchmark performance
0 likes · 4 min read


DevOpsFeb 25, 2025 · Artificial Intelligence
Claude 3.7 Sonnet, Anthropic's new hybrid reasoning model, introduces dual thinking modes, token‑based thinking budget control, unchanged pricing, and the Claude Code tool that automates lengthy coding tasks, while achieving record GPQA scores, superior video‑game testing results, and reduced unnecessary refusals on harmful requests.
AI ModelClaudeCoding tool
0 likes · 7 min read

GPT‑4, the new multimodal AI model, can process images and text, generate code and natural language, achieve human‑level scores on standardized exams, handle up to 32 K tokens, and demonstrates advanced reasoning, while OpenAI emphasizes its safety improvements and current limitations as a still‑emerging technology.
AI safetyGPT-4benchmark performance
0 likes · 6 min read
