Moonshot AI Releases Kimi K3 Open Weights: A 1.56 TB Sparse MoE Model on Hugging Face

·1 views

Moonshot AI last week published the Kimi K3 model weights on Hugging Face, making available a 2.8-trillion-parameter sparse mixture-of-experts model compressed in MXFP4 format. The full checkpoint spans roughly 1.56 TB across 96 safetensors shards, accompanied by configuration, tokenizer, and evaluation metadata files. Serving the model requires compatible software such as vLLM or SGLang, with support varying depending on the user's GPU generation. The model is released under a custom, non-OSI license that permits commercial use but imposes revenue-tier conditions and lists prohibited use cases. Developers are advised to pin a specific repository revision before downloading to avoid inconsistencies if the upstream files are updated.

Read the full story at DEV Community

This is an AI-generated summary. ShortSingh links to the original source for the complete article.

Discussion (0)

Developer Documents Journey to Build an AI Operating Layer From First Principles

A software engineer has launched a public engineering journal chronicling the development of a project called the AI Operating Layer. Rather than rushing into implementation, the developer spent months studying AI tools, frameworks, and agent platforms to avoid solving the wrong problem. The project aims to unify models, memory, orchestration, tools, and execution into a single interconnected ecosystem rather than treating them as separate components. The developer originally conceived it as a synthesis engine but has since reframed it as a living knowledge ecosystem capable of continuous learning and growth. The journal is intended to document not just successes but also wrong turns, redesigns, and evolving ideas throughout the build process.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Registry tracking 218 AI and SaaS tools finds 18 dead or sunset, reveals key patterns

A developer launched Verdict, a registry that monitors the live status of 218 AI and SaaS tools, and has so far confirmed 18 deaths or sunsets with dated, sourced records. Analysis of these closures revealed recurring patterns, including acqui-hires where the acquiring company shuts down the original product despite promising to continue its mission. Tools like Xembly gave users as little as two days' notice before shutting down, highlighting the risk of depending on venture-backed software without a fast export plan. A split observed with OpenAI's Sora — where the app was killed but the API lives on until September 2026 — suggests that app shutdowns can serve as early warnings for developers building on those platforms. The project also found that closures tend to cluster by category, and that most "best tools" listicles continue recommending products that no longer exist, underscoring the need for regularly verified recommendations.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Researcher Discovers High-Severity AI Security Flaw in Khan Academy's VDP

A security researcher identified a high-severity vulnerability in Khan Academy's Vulnerability Disclosure Program (VDP) related to its AI systems. The flaw was uncovered using recon techniques and AI prompt testing tools. Khan Academy operates a VDP to allow ethical hackers to responsibly report security issues in its platforms. The researcher published their findings on the GaleOps Blog, detailing the methods and tools used during the discovery process.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

OpenAI Launches Codex Security in Research Preview for Enterprise Customers

OpenAI has introduced Codex Security as a research-preview feature within its Codex web platform, available to ChatGPT Pro, Enterprise, Business, and Edu subscribers. The tool is designed to build threat models based on project context, validate potential security findings in sandboxed environments, and surface proposed code patches for human review. By combining codebase context with automated validation, the system aims to make security findings more actionable and reduce the triage burden on engineering teams. A human reviewer remains part of the decision loop for any proposed changes, reflecting OpenAI's framing of the tool as a review-and-remediation workflow rather than a standalone vulnerability scanner. Codex Security is a cloud-hosted, enterprise-oriented capability and is not currently available as a general-purpose or open-source security tool for all developers.

0 comments Read more at DEV Community