A developer in Kenya benchmarked 16 coding models across six languages — English, Chinese, French, Spanish, Swahili and the low-resource Luhya language Maragoli — finding that 11 of 16 scored 100% on …
A developer benchmarked 10 AI models as judges on Kaggle, planting errors in their own prior answers to test self-preference, and found that only two small OpenAI models graded their own mistakes more…
A developer built a 48-scenario benchmark, verification_honesty, that tests whether coding agents honestly report the verification status of their own work by reading realistic session logs and fillin…
A Tamil-speaking developer built a 60-question benchmark comparing 13 AI models on English, Tamil script, and romanized Tanglish to measure what he calls the "script penalty." Google's Gemini 3.5 and …
A growing body of 2026 research shows that an AI agent's harness — its loop, context management, tool interface and recovery logic — drives benchmark performance far more than the underlying model wei…
MiniMax released Mini-Agent, a free terminal-based CLI coding agent that developers install with one command and connect to an API key from platform.minimax.io, offering unlimited usage without the ca…
Z.ai (Zhipu) has deprecated GLM-4.7, with the model set to retire on 2026-12-31 and be replaced by GLM-5. The glm-4.7 model, which offers a 200k context window and 66k max output, is flagged as retiri…
A developer built the AI Security Stress-Test Benchmark on Kaggle, a 12-task evaluation spanning code vulnerabilities, cloud misconfigurations, and adversarial jailbreaks, and ran six frontier models …
A developer built Cyber Autopsy, a Kaggle benchmark that tests whether AI models can reconstruct reported cyber incidents from partial evidence while citing support for each claim and flagging what re…
A developer built "Blog vs Bytecode," a Kaggle benchmark of 28 data-science code snippets paired with blog-style claims, half of which hide real methodological errors such as scaler leakage or shuffle…
Developer "&." released an early-prototype webpage at ampdot.mesh.host that generates a token-space font by combining a base font with a user-selected tokenizer, giving every token equal width so AI o…
Zhipu AI's ChatGLM assistant runs on the open-weight GLM model family, with GLM-4.5 and the GLM-5 series published under the MIT license by the zai-org GitHub organisation and on Hugging Face. Zhipu A…
Z.ai revealed that the anonymous OpenRouter model ox-alpha was actually GLM-5.3-Flash, the smaller version of its GLM-5 model. Before the official announcement, developers used tokenizer fingerprintin…
A developer's analysis of pre-tokenization regexes shows that GPT-2's word-splitting pattern omitted Unicode's Mark category, a bug inherited by GPT-4, Llama 3, Qwen 3, and GLM-4/5, forcing BPE to tok…
Z.ai announced on September 1 that every current GLM Coding Plan subscriber will receive a Reset Card that refills both the rolling five-hour allowance and the weekly quota, with redemption requiring …
IBM Research's ALT K-Evolve framework shows that the optimal amount of agentic memory varies by model capability, with strong models like DeepSeek-V3.2 (671B MoE) gaining +9.5 percentage points in tas…
A June 2026 study by LightSpeed and Tencent found that four frontier models—Claude Sonnet 4.5, DeepSeek-V4-Pro, GLM-5, and Gemini-3-Flash—misjudged their total context size with median relative error …
Vernor Vinge, a computer scientist and science fiction author, predicted in his 1993 essay that the Singularity—a point where machines surpass human intelligence and accelerate progress beyond human c…
Code Arena's WebDev leaderboard, updated August 6, 2026, shows the Elo gap between proprietary and open-weight AI coding models has narrowed from roughly 150 points to just 11, with Anthropic's Claude…
TileRT's persistent engine on NVIDIA GPUs achieves up to 500 tokens/s/user on the InferenceX GLM5 FP8 744B benchmark on a single B200 decode server, approximately 3× faster than GB300 NVL72 running tr…