26B 参数但活跃只有 5.4B,MacBook 就能跑,Apache 2.0 可商用,还带 1M 上下文。想在自己设备上部署开源模型、或对比同量级 MoE 效率的开发者,可以直接下权重试试。
#Inkling 2 Mini
Thinking Machines Lab 这次的 Inkling 2 Mini 只有 26B 总参数、激活 3B,普通本地机器就能跑。transformers、llama.cpp 和 MLX 都已同步支持,Mac 用户用 MLX 直接上手,想在本地跑编程助手的可以试试和 Qwen3 对比一下。
Mira Murati 团队这次的 Inkling 2 Mini 有意思:中间层每个 token 跑三遍,用更少 token 就能完成推理。20GB 显存笔记本能跑,DeepSWE 48.2% 拿下百亿级以下开源模型第一。想在本地折腾编程推理的可以试试。
The monitorability claim is the real story: looping compresses reasoning into an inspectable process, offering a different answer to AI-safety concerns than simply reading longer chains of thought. Developers running inference on local hardware should try Inkling 2 Mini, since it removes the need for API access and data leaves the device. Compared with cloud reasoning models from OpenAI or DeepSeek, it trades raw capability for transparency and offline availability.
A looped architecture letting a 26B model carry agentic workloads is the notable detail, since weight reuse rather than scale is doing the heavy lifting. Developers who want on-device agents without cloud APIs should benchmark it against other small local models in their own workflows. The AA Index score gives a first concrete comparison point, though independent laptop performance tests will matter more than vendor claims.