Tag archive
한국 AI 스타트업 VIDRAFT가 6.59B 파라미터 MoE 모델 Aether-7B-5Attn을 Apache-2.0으로 공개. 가중치뿐 아니라 학습 데이터·코드·로그·체크포인트까지 전면 공개한 진정한 오픈소스 기반 모델.
한국 스타트업 VIDRAFT가 70억 파라미터 기반 모델 Aether-7B-5Attn을 Apache-2.0으로 공개. 가중치뿐 아니라 학습 코드, 하이퍼파라미터, 로그, 중간 체크포인트까지 완전 투명 공개. 연구·상업 활용 모두 가능.
한국 AI 스타트업 VIDRAFT가 70억 파라미터 파운데이션 모델 Aether-7B-5Attn의 가중치·학습 코드·학습 데이터를 전면 공개했습니다. ML 엔지니어라면 주목해야 할 오픈사이언스 릴리즈.
Originally published on andrew.ooo — visit the original for any updates, code snippets that aged...
A practical 2026 guide to fine-tuning open-source LLMs with LoRA and QLoRA using Unsloth + Gemma 4 — including GPU requirements, hyperparameter defaults, evaluation setup, and when to just prompt instead.
Zhipu AI's 753B open-weight GLM-5.2 is the highest-ranking open-source model on lmarena.ai, challenging Claude Fable 5 across WebDev and Agent benchmarks — and it's already runnable locally via Ollama.
MiniMax M3 drops 428B parameters, a 1M-token context window, and open-source weights for free — I tested it against Claude Code on real coding tasks and found surprising results.
On June 13, 2026, the US government cracked down on Anthropic's Claude Fable 5. Hours later, China's ZhipuAI open-sourced GLM 5.2 under MIT license — with a 1M context window and frontier-grade coding scores. This is what happened, why it matters, and how to use it today.
The open-source coding LLM leaderboard looked completely different in April than it does today. MiniMax M3 just shipped June 1st. GLM-5.1 landed in April with an 8-hour autonomous execution claim nobody expected. Here's the real picture as of June 2026.
DeepSeek V4-Pro is a 1.6T MIT-licensed model with 80.6% SWE-bench. This guide covers API setup, pricing vs GPT-5.5, and self-hosting options.
Qwen3.6-27B delivers flagship-level agentic coding in 27B dense parameters, outperforming 397B MoE models. Apache 2.0 local deployment guide.
Arcee Trinity Large Thinking is a 400B Apache 2.0 sparse MoE model built for long-horizon agents. API, self-hosting, benchmarks, and integration guide.