Run 35B MoE models in ~2.5 GB of RAM. Weights stream from SSD; a trained router predicts loads a step ahead. Mac, iPhone, Android, Windows supported
Python
README: name base models explicitly in quality tables
linyubupaa615d1387ea0
diff --git a/README.md b/README.mdindex 8052d58..097df68 100644--- a/README.md+++ b/README.md@@ -215,7 +215,7 @@ original fp16 base models — the loss is small: **3.9 points on average foredge0-35b, 2.8 for edge0-8b** (MMLU-Pro is even above the base). All scoresself-run, max 100:-| Benchmark | edge0-35b (int4) | Base fp16 | edge0-8b (int4) | Base fp16 |+| Benchmark | edge0-35b (int4) | Qwen3.5-MoE 35B-A3B (fp16) | edge0-8b (int4) | Ling 3.0 tiny (fp16) ||---|---:|---:|---:|---:|| AIME 2026 | 86.6 | 92.7 | 63.3 | 73.3 || HumanEval | 90.9 | 95.1 | 91.5 | 92.7 |diff --git a/README_zh.md b/README_zh.mdindex 02d98a1..14b567a 100644--- a/README_zh.md+++ b/README_zh.md@@ -180,7 +180,7 @@ engine.close() # 释放 mmap / 专家缓存3.9 分、edge0-8b 落后 2.8 分**(MMLU-Pro 甚至反超基座)。均为自测,满分 100:-| 评测项 | edge0-35b (int4) | 基座 fp16 | edge0-8b (int4) | 基座 fp16 |+| 评测项 | edge0-35b (int4) | Qwen3.5-MoE 35B-A3B(fp16) | edge0-8b (int4) | Ling 3.0 tiny(fp16) ||---|---:|---:|---:|---:|| AIME 2026 | 86.6 | 92.7 | 63.3 | 73.3 || HumanEval | 90.9 | 95.1 | 91.5 | 92.7 |