sddqwen35a3b_v01 is an open-weight model from Max Sun. Its published files total 960.6 KB.
在 8 GB RAM 的 Intel 機器上真的跑得動 Qwen3.6-35B-A3B（22.1 GB / Q4KM）： 冷專家留在 SSD，RAM 只放有限容量的 expert cache，不用 swap 冒充 RAM。 - 引擎基礎：llama.cpp，用 patch 改（不重寫引擎）。唯一的進入點是 llamaserver.sh。 - 核心實作：cpp/sdqpager.cpp（RAM arena + 大小兩種槽位 + pin + 淘汰 + 指標） ＋ cpp/sdqmoe.cpp（自訂 MoE op）＋…
Parameters—
Context—
Weights960.6 KB
License—
AccessOpen weights
Monthly Downloads—
Model Card
在 8 GB RAM 的 Intel 機器上真的跑得動 Qwen3.6-35B-A3B（22.1 GB / Q4KM）： 冷專家留在 SSD，RAM 只放有限容量的 expert cache，不用 swap 冒充 RAM。 - 引擎基礎：llama.cpp，用 patch 改（不重寫引擎）。唯一的進入點是 llamaserver.sh。 - 核心實作：cpp/sdqpager.cpp（RAM arena + 大小兩種槽位 + pin + 淘汰 + 指標） ＋ cpp/sdqmoe.cpp（自訂 MoE op）＋ cpp/sdqbuft.cpp（讓 18.2 GiB 專家權重不進 RAM） - 進度真相來源：stage.txt（每階段更新，並上傳 HF） 1. git clone llama.cpp 並 checkout 到 patch 對應的 commit 2. 抓 patches/.patch（旁邊沒有就從 HelloSun/sddqwen35a3bv01 抓回來） 3. git apply 套 patch（先 --check，套不上就停下來報錯） 4. cmake 編譯 llama-server 6. 啟動 llama-server 需求只有 git、cmake、curl、python3 和 C++ toolchain。 unsloth/Qwen3.6-35B-A3B-GGUF 的 Qwen3.6-35B-A3B-UD-Q4KM.gguf（22,134,528,992 bytes）， 永遠只做 SSD 上的 file-backed mmap。…
Excerpt from the card by Max Sun.
Identity and Version
- Repository
- HelloSun/sddqwen35a3b_v01
- Publisher
- Max Sun
- Task
- Not stated by the source
- Modality
- Other
- Library
- Not stated by the source
- Parameters
- Not stated by the source
- Languages
- Not stated by the source
- Revision
- a5654497d993dced0b0769aefcb69acc8961f5a7
- First published
- 2026-10-05
- Last updated
- 2026-10-06
Files and Weights
97 files, 960.6 KB in total.
Configuration39 files · 242.9 KB
Documentation28 files · 122.7 KB
Other28 files · 593.3 KB
Repository2 files · 1.7 KB
Every file
| File | Type | Size | SHA-256 |
|---|---|---|---|
| bench/expert_index.json | Configuration | 33.2 KB | — |
| bench/results/bench_first.json | Configuration | 6.8 KB | — |
| bench/results/bench_pipeline.json | Configuration | 6.8 KB | — |
| bench/results/bench_s2.json | Configuration | 6.5 KB | — |
| bench/results/bench_s4_modec.json | Configuration | 6.4 KB | — |
| bench/results/exp_final_ab.json | Configuration | 2.7 KB | — |
| bench/results/exp_rest_modec.json | Configuration | 4.1 KB | — |
| bench/results/exp_s5_io.json | Configuration | 3.1 KB | — |
| bench/results/exp_s5_io2.json | Configuration | 4.0 KB | — |
| bench/results/exp_s6_slot_ab.json | Configuration | 1.1 KB | — |
| bench/results/exp_s7_pf.json | Configuration | 2.1 KB | — |
| bench/results/exp_s7_pf_confirm.json | Configuration | 1.1 KB | — |
| bench/results/exp_s8_release.json | Configuration | 1.1 KB | — |
| bench/results/exp_s9_final.json | Configuration | 1.1 KB | — |
| bench/results/exp_s9_final2.json | Configuration | 1.1 KB | — |
| bench/results/exp_s9_final3.json | Configuration | 1.3 KB | — |
| bench/results/exp_s9_racefix.json | Configuration | 1.2 KB | — |
| bench/results/exp_smoke.json | Configuration | 884 B | — |
| bench/results/memory_ab_same-machine-2g-system.json | Configuration | 7.2 KB | — |
| bench/results/memory_levels_4-8-12g-budget.json | Configuration | 4.6 KB | — |
| sdq/__init__.py | Configuration | — | |
| sdq/bench.py | Configuration | 7.0 KB | — |
| sdq/config.py | Configuration | 4.5 KB | — |
| sdq/detect.py | Configuration | 12.1 KB | — |
| sdq/experiments.py | Configuration | 17.0 KB | — |
| sdq/hf_sync.py | Configuration | 7.2 KB | — |
| sdq/memory_ab.py | Configuration | 17.8 KB | — |
| sdq/memory_levels.py | Configuration | 9.8 KB | — |
| sdq/verify.py | Configuration | 14.6 KB | — |
| tools/determinism_check.py | Configuration | 5.3 KB | — |
| tools/drop_model_cache.py | Configuration | 1.6 KB | — |
| tools/gguf_index.py | Configuration | 10.1 KB | — |
| tools/hot_expert_compress_bench.py | Configuration | 10.5 KB | — |
| tools/io_probe.py | Configuration | 4.3 KB | — |
| tools/io_probe_ceiling.py | Configuration | 4.1 KB | — |
| tools/moe_layer_check.py | Configuration | 3.0 KB | — |
| tools/moe_ref.py | Configuration | 9.2 KB | — |
| tools/per_layer_report.py | Configuration | 3.4 KB | — |
| tools/route_overlap.py | Configuration | 5.2 KB | — |
| README.md | Documentation | 18.5 KB | — |
| bench/results/exp_final_ab.md | Documentation | 715 B | — |
| bench/results/exp_rest_modec.md | Documentation | 939 B | — |
| bench/results/exp_s5_io.md | Documentation | 875 B | — |
| bench/results/exp_s5_io2.md | Documentation | 1.1 KB | — |
| bench/results/exp_s6_slot_ab.md | Documentation | 398 B | — |
| bench/results/exp_s7_pf.md | Documentation | 584 B | — |
| bench/results/exp_s7_pf_confirm.md | Documentation | 383 B | — |
| bench/results/exp_s8_release.md | Documentation | 405 B | — |
| bench/results/exp_s9_final.md | Documentation | 405 B | — |
| bench/results/exp_s9_final2.md | Documentation | 424 B | — |
| bench/results/exp_s9_final3.md | Documentation | 405 B | — |
| bench/results/exp_s9_racefix.md | Documentation | 393 B | — |
| bench/results/exp_smoke.md | Documentation | 382 B | — |
| bench/results/memory_ab_same-machine-2g-system.md | Documentation | 1.6 KB | — |
| bench/results/memory_levels_4-8-12g-budget.md | Documentation | 1.2 KB | — |
| docs/01_moe_path_analysis.md | Documentation | 5.9 KB | — |
| docs/02_design.md | Documentation | 6.3 KB | — |
| docs/03_results.md | Documentation | 5.5 KB | — |
| docs/04_experiments.md | Documentation | 13.9 KB | — |
| docs/05_optimization.md | Documentation | 11.0 KB | — |
| docs/06_slots.md | Documentation | 7.3 KB | — |
| docs/07_prefetch.md | Documentation | 6.4 KB | — |
| docs/08_verify_gaps.md | Documentation | 6.5 KB | — |
| docs/09_determinism.md | Documentation | 7.2 KB | — |
| docs/10_phase0_measurement.md | Documentation | 13.2 KB | — |
| docs/hot_expert_lossless_compression_2026-10-06.md | Documentation | 5.6 KB | — |
| docs/performance_report_2026-10-06.md | Documentation | 5.5 KB | — |
| bench/results/exp_s5_io_final.txt | Other | 463 B | — |
| bench/results/exp_s5_io_run1.log | Other | 1.8 KB | — |
| bench/results/exp_s6_slot_ab.log | Other | 971 B | — |
| bench/results/exp_s7_pf.log | Other | 1.3 KB | — |
| bench/results/route_decode48.log | Other | 89.5 KB | — |
| bench/results/s9_det2.log | Other | 418 B | — |
| bench/results/s9_final3.log | Other | 1.0 KB | — |
| bench/results/s9_race.log | Other | 1.4 KB | — |
| bench/results/s9_racefix.log | Other | 2.4 KB | — |
| bench/results/stats_cold.jsonl | Other | 1.0 KB | — |
| bench/results/stats_hot0.jsonl | Other | 1.0 KB | — |
| bench/results/stats_hot1.jsonl | Other | 1.0 KB | — |
| bench/results/verify_s6.log | Other | 751 B | — |
| bench/results/verify_s7.log | Other | 1.7 KB | — |
| bench/results/verify_s8_full.log | Other | 2.2 KB | — |
| bench/results/verify_s8_pfstress.log | Other | 461 B | — |
| cpp/sdq_buft.cpp | Other | 8.0 KB | — |
| cpp/sdq_cli.cpp | Other | 31.8 KB | — |
| cpp/sdq_moe.cpp | Other | 32.7 KB | — |
| cpp/sdq_pager.cpp | Other | 79.9 KB | — |
| cpp/sdq_pager.h | Other | 19.9 KB | — |
| llama_server.sh | Other | 19.4 KB | — |
| patches/0001-sdq-expert-pager.patch | Other | 176.4 KB | — |
| patches/0002-sdq-server-hook.patch | Other | 3.2 KB | — |
| patches/0003-sdq-measured-defaults.patch | Other | 4.0 KB | — |
| patches/0004-sdq-observability.patch | Other | 9.9 KB | — |
| patches/0005-sdq-server-integration.patch | Other | 6.0 KB | — |
| stage.txt | Other | 94.7 KB | — |
| .gitattributes | Repository | 1.5 KB | — |
| .gitignore | Repository | 162 B | — |
License and Download
- License
- Not stated by the source
- Access
- Open weights, no gate
Download from Max Sun
Released by Max Sun through its official repository on Hugging Face.