Top-5 Best Value #LLM Models of the Day at UTC-04
| Model | #AAII | Price | | DeepSeek V4 Flash 0731 | 52 | $0.04 | | GLM 5.3 Flash | 58 | $0.09 | | Gemini 3.7 Flash | 56 | $0.29 | | GPT-5.6 Luna | 52 | $0.18 | | Qwen3.8 2.4T A95B | 58 | $2.06 |
I deployed "Henan Nier" locally on MacBook!
This issue adds the last building block to my digital human "Henan Ni'er": I ran the digital human completely locally on a MacBook with 16GB of memory, and she spoke in about 1.3 seconds after she finished speaking. Plan details: • Digital human: 3D image in VRM format, only rendering, not generating, and does not occupy any reasoning resources • Rendering: three.js +…
Gaming notebooks can now directly run cutting-edge large models, with official weights, without extreme quantification.
Qwen3.6 35B, 4060 notebook with 8G video memory, 39 tok/s. DeepSeek-V4-Flash 284B, 5090 desktop, 22-25 tok/s. GLM-5.2 753B, one PRO 6000, 15 tok/s. Claude Code and Codex are direct local prostitutes, and the electricity bills are cheaper than API.
MoE compute is sparse. MoE memory is not.
A token activates 2 of Mixtral's 8 experts (25% of FFN weights), or 8 of DeepSeek-V3's 256 (~5%). But every expert must live SOMEWHERE, and no single GPU holds them all. Expert parallelism inverts the usual rule: experts stay put, spread
You have reached the end of the archive
All of deepseek