DeepSeek-V4-Pro defeated Llama-4-Maverick-17B-128E-Instruct-FP8 with an aggregate score of 101.3 to 86.2, winning 9 of 12 tasks with 98% confidence in a head-to-head benchmark conducted by an unnamed …
A new framework called DARWIN uses a genetic algorithm to evolve jailbreak prompts for large language models, achieving nearly 100% success rates on DeepSeek-V4-Pro and over 90% on GPT-5.5. The DARWIN…
Google's Gemini 3.6 Flash defeated DeepSeek-V4-Pro in a head-to-head text task evaluation, winning 5 tasks to 2 with 86% confidence and an overall score of 111.7 to 98.8, according to tests run by an …
DeepSeek-V4-Pro defeated Sarvam M 113.0 to 44.4 in a 12-task benchmark, achieving a 12-0 sweep with 100% confidence. The test, judged twice by gpt-5.4 to cancel position bias, found DeepSeek-V4-Pro re…
Gpt-oss-120b defeated DeepSeek-V4-Pro 106.3 to 93.7 across 12 text tasks, winning 10 of 12 matchups with 99% confidence, according to a head-to-head benchmark scored by gpt-5.4. The open-source model …
An independent test of DeepSeek-V4-Pro found sharply different behavior depending on whether prompts contained cyber or biology content, raising new questions about whether some requests were being an…
EvoClawBench, a new benchmark testing whether AI agents can learn reusable skills from experience, shows mixed results across 100 tasks and 502 sub-problems. Nanobot's GPT-5.4 model consistently excee…
Muse Spark 1.1 defeated DeepSeek-V4-Pro 11-1 in a head-to-head benchmark of 12 text tasks, scoring 110.8 to 84.3 with 100% confidence. The model outperformed in localization, proofreading, JSON extrac…
DeepReinforce released Ornith-1.0, an open-source family of coding models that learn to build their own task scaffolds, moving intelligence from the surrounding harness into the model itself. The mode…
Inferize announced DeepSeek-V4-Pro, claiming it can serve the model in 20 seconds with highly optimized, elastic AI inference. The company is building fast, efficient LLM serving that scales with dema…
DeepReinforce open-sourced Ornith-1.0, a family of self-improving coding models ranging from 9B to 397B parameters, which learn to generate their own task-specific scaffolds during reinforcement learn…
AMD shipped ATOM + ATOMesh, a ROCm-native LLM serving stack for Instinct GPUs that implements prefill/decode disaggregation, splitting the two inference phases onto separate GPU pools to optimize for …
DeepSeek AI released preview versions of its DeepSeek-V4 series, including two Mixture-of-Experts language models with up to 1.6 trillion parameters and support for one-million-token contexts. The mod…
A developer reports that the downloadsAllTime field for multiple public models on Hugging Face Hub is decreasing day-over-day, contrary to expectations for a cumulative count. The observed drops range…
OpenRouter's Fusion model outperformed DeepSeek-V4-Pro in a head-to-head coding test, delivering a complete, compilable Go rate limit parser with correct edge cases while DeepSeek V4 Pro made basic er…
NVIDIA achieved leading agentic coding performance on the first agentic AI benchmark, AA-AgentPerf, delivering up to 20x better performance than previous generations. The benchmark, created by Artific…
DeepSeek-V4 shifts the focus from simply offering a million-token context window to making that window practically usable through efficient attention mechanisms. Its architecture employs hybrid compre…
DeepSeek-V4 introduces a new architecture using hybrid attention mechanisms—Compressed Sparse Attention (CSA) and Heavily Compressed Attention (HCA)—to drastically reduce the computational cost and me…