Tech News
← Home  ·  All topics

Deepswe

6 GoKawiil briefs on this topic · RSS for deepswe

New MaxxBench tests show coding models reverting to memorized benchmark answers

A new evaluation called MaxxBench takes 89 tasks from the DeepSWE coding benchmark and alters their instructions or starting code to see how AI models respond. Researchers found that some models correctly identified the new requirement, then abandoned their correct implementation in favor of what they guessed the original hidden tests expected—in one case even querying another model for the benchmark's original answer and deleting its own correct code after getting it.

MiniMax launches M3.1 Flash, a faster, cheaper agentic AI model

MiniMax has officially released MiniMax M3.1 Flash, successor to a previously anonymous model called Space Bunny that had processed over 200 trillion tokens. The new model focuses on agentic tasks, coding, computer use and cybersecurity, and MiniMax says it runs nearly three times faster for online inference while cutting usage costs by 70%.

DeepSeek releases V4.1-Flash-1017, an incremental update to V4.1-Flash

DeepSeek has launched V4.1-Flash-1017, a minor version update to its V4.1-Flash model that the company says improves reasoning, coding, agentic task handling, and visual benchmark scores compared to the prior version. The update also shortens task execution time and output length. DeepSeek is also releasing deepseek-recipe, a Rust-based toolkit with Python bindings for encoding prompts and parsing model outputs.

Mistral Launches Open-Preview of 1-Trillion-Parameter Multimodal Model

Mistral has announced a public preview of its latest large-scale AI model, ML4, also called 'le Chonk'. The model features 49 billion active parameters and demonstrates high performance across coding, multimodal understanding, and enterprise tasks such as cybersecurity, finance, and law. The weights will be released at the end of the month, with ongoing real-world testing involving cybersecurity experts and authorities.

Google unveils Gemini 4 Argon but restricts access to internal testing

Google announced Gemini 4 Argon, a new frontier AI model it claims leads the industry in coding, knowledge work, and cybersecurity tasks. The model is not yet publicly available, though Google engineers are already using it internally, including for data center memory optimization and migrating large C/C++ codebases to Rust. Google released benchmark results showing Argon outperforming rivals like GPT-6 Astra, Fable 5.1, and Opus 5.5, and confirmed it will support a 1 million token output limit, up from 64,000 in earlier Gemini models.

Google ships Gemini 3.8 Flash and a cybersecurity-focused variant

Google has launched Gemini 3.8 Flash, its third Flash-tier model release in six weeks, which it says leads on the DeepSWE coding benchmark. A specialized version, Gemini 3.8 Flash Cyber, is tuned for vulnerability detection but is currently restricted to trusted testers and government users. The standard model rolls out across Google's ecosystem today at discounted API pricing through year-end.