Local benchmark: Qwen 3.8 and DeepSeek 4 show significant quality jump on Mac
olcan · x · 2026-08-17
User tested Qwen 3.8 27B and DeepSeek 4 Flash locally on a Mac. Both models represent a significant jump in quality, remaining very usable at 20-30 tps even for longer contexts up to 100k tokens.
More from Models
- Users claim Qwen3.5 9B quantized model outperforms ChatGPT-4o with only 7GB VRAM — ML-Future · 2026-08-17
- Petition calls for mandatory quantization labels in model评测 posts — Su1tz · 2026-08-17
- OpenAI reportedly starts teasing next model 'Astra' for potential release — haider1 · 2026-08-17
- Dev on OpenAI 1M context: Seamless compaction is the real win — eyishazyer · 2026-08-17
- Users report Qwen models "overthink" simple responses, exhausting context — malliktwts · 2026-08-17
- Sarvam AI Unveils 105B Voice Model and Full-Stack Agent Solutions — AashaySachdeva · 2026-08-17