The article benchmarks three domestic multimodal large models—Step 3.7 Flash, Qwen 3.6‑flash, and MiniMax M3—across two production‑oriented scenarios, measuring quality, latency, and token cost, and concludes that Step 3.7 Flash consistently offers the best speed‑cost trade‑off while maintaining reliable output.
The vLLM 0.24.0 update adds MiniMax‑M3, DeepSeek‑V4, DiffusionGemma support, a Streaming Parser Engine, and a new device_ids parameter, delivering faster inference, lower memory use, and broader hardware compatibility for large‑model deployments.
A previously unknown Chinese AI security group called MopMonk, operating without a website or corporate backing, posted a GitHub report that achieved a 73.1% vulnerability‑exploitation success rate, ranked seventh globally in the UC Berkeley‑run CyberGym benchmark, and demonstrated novel memory‑based multi‑agent techniques that signal China’s rising AI security prowess.
The article benchmarks three Chinese multimodal large models—Step 3.7 Flash, MiniMax M3, and Qwen 3.6‑flash—across two real‑world tasks, measuring output quality, API latency, and token cost, and concludes that Step 3.7 Flash consistently offers the best speed‑cost trade‑off for production use.
The author compares four Chinese large‑language models—GLM 5.2, Kimi 2.7, DeepSeek V4 and MiniMax M3—detailing their strengths, pricing, and ideal use‑cases for writing, coding, multimodal processing and high‑throughput batch tasks, and shares personal trust insights.
MiniMax unveiled its open‑source M3 model, delivering 1 million‑token context, 59 % SWE‑Bench Pro accuracy that outperforms GPT‑5.5 and Gemini 3.1 Pro, native multimodal desktop interaction, and a 9‑15× speed boost via MiniMax Sparse Attention, with pricing as low as $20 per month.
The article walks through a hands‑on assessment of MiniMax M3, using a real Redis SCAN outage to test fault diagnosis, then reproducing the SCAN algorithm in Go and finally constructing a lightweight Redis monitoring panel, while comparing its performance to GLM‑5.1.
The author evaluates MiniMax M3 by reproducing the 2021 TransUNet medical image segmentation paper, troubleshooting data and training issues, achieving near‑paper Dice scores, and then engineering a full‑stack React‑FastAPI application to demonstrate the model’s practical capabilities and cost.
MiniMax M3 claims to be the first open‑source LLM that simultaneously delivers top‑tier coding/agentic ability, a 1‑million‑token context window, and native multimodal understanding, and our benchmarks on coding suites, long‑context efficiency, and multimodal tasks confirm it exceeds expectations.
MiniMax M3, a domestically developed LLM, combines a new sparse‑attention MSA architecture, native multimodal support, and million‑token context to match or surpass top closed‑source models in programming and agent benchmarks, while achieving a 9.4× speedup on FP8 GEMM and preparing for open‑source release.
The article reveals that xAI’s upcoming Grok-5 (Grok V9-Medium) will feature a 1.5-trillion-parameter model trained with extensive Cursor programming data, while MiniMax-M3 introduces a new sparse-attention architecture that boosts pre-fill speed by 9.7× and decode speed by 15.6×, highlighting a strategic partnership between SpaceX, Cursor, and xAI.
