@zhixianio: After receiving the new machine, I began an 'ascetic' practice of forcing myself to use local models for common tasks. I thought it would be painful, but both speed and quality greatly exceeded my expectations: Model: Qwen3.6-35B-A3B-oQ6-fp16-mtp, Running: oMLX, with N…
Summary
The author uses the Qwen3.6-35B-A3B model and oMLX tool on the new local machine for daily tasks, finding that both speed and quality far exceed expectations, even outperforming remote LLMs in PA and coding scenarios, demonstrating a significant improvement in on-device AI capabilities.
View Cached Full Text
Cached at: 06/03/26, 07:46 AM
After the new machine arrived these past two days, I began an ascetic-like practice of forcing myself to use local models to complete common tasks.
Expected it to be extremely painful, but to my surprise, both the speed and quality far exceeded my expectations:
Model: Qwen3.6-35B-A3B-oQ6-fp16-mtp Running: oMLX, Native MTP enabled, Thinking disabled, CTX 128K
These days, whether in the PA scenario (OpenClaw) or the Coding scenario (PI-Mono), the response speed is faster than remote LLMs, the intelligence is on-point, and it can even help me think and do one more step. The native multimodal capabilities under PA are even more satisfying to use than DSV4 Pro.
I never expected that the capabilities of on-device models have reached this level.
Similar Articles
@RookieRicardoR: Domestic models break through again, matching top models like Claude 4.6 and Gemini 3.1 Pro. Just tested Qwen3.7-Max, sharing some real thoughts. Last night I topped up as soon as the API went live and chose three tasks (see video) to test Qwen3.7-Max's frontend capabilities…
The user tested Qwen3.7-Max and believes it matches top models like Claude 4.6 and Gemini 3.1 Pro in frontend, computing power, and Agent capabilities. Its reasoning ability has significantly improved, and with monthly iteration speed, it has become a first-tier domestic model.
@wquguru: Local deployment of models can't do without some acceleration tricks. Simply put, you can switch to models like Turbo; …
作者系统整理了本地部署模型的加速方案，包括在 macOS 上结合 Turbo、MLX、VAE Q8 和前缀 KV cache 让 Qwen-Image 2.1 提速近 5 倍，以及用 Uzu 的投机解码把 Qwen3.5-9B 从 22 tok/s 提升到 92 tok/s、用 TensorRT 让 Qwen 在 RTX 5090 上比 llama.cpp 快 3 倍等案例。
@lilililiMozi: About the fact that I am clearly the Qwen3.8 27B model, yet underwent specialized fine-tuning with 3.9 million lines of…
发布者声称该模型基于 Qwen3.8 27B，使用 390 万行代码数据进行专项微训，融合高效思考机制压缩推理长度，并经由 Opus5.5、GPT6Astra、Grok4.7、DSV4Pro 蒸馏对齐，配备 K3 kernel，再通过 SFT、RLOO 强化学习与 MTP 多词元预测三重训练，同时集成 DFlash2 注意力加速。
@yibie: Using Local Models as Primary Coding Tools: A Practical Report from Mid-2026 There was a post on Hacker News with a straightforward title: "Is anyone using local models as their primary coding tool?" 197 comments, incredibly dense with information. A dozen real users discussed their daily configurations, pitfalls they encountered, and why they still choose local models even though they know they're not as good as...
This article summarizes practical experiences from a Hacker News discussion about using local models (mainly Qwen 3.6 35B-A3B) as primary coding tools, including configurations, effectiveness (approximately 50-75% of frontier models), key techniques (such as preserve_thinking), and different user positions.
@noctus91: I recently switched from Qwen 3.5 9B to LFM2.5-8B-A1B by @liquidai, and it's quickly become my default local model in H…
A user shares their positive experience switching from Qwen 3.5 9B to Liquid AI's new LFM2.5-8B-A1B model, praising its speed and reliability for agentic tasks while noting coding remains a weakness. The model is an 8B MoE with 1.5B active parameters and 128K context, optimized for devices and server-side use.