Skip to page content
SAVRN
Search Plan a project

Open-weight model · Text generation

qwen35a3b-moe-expert-paging-q4

by Max Sun HelloSun/qwen35a3b-moe-expert-paging-q4

qwen35a3b-moe-expert-paging-q4 is an open-weight model for text generation from Max Sun, released under Apache License 2.0. Its published files total 841.8 KB.

EPQ(Expert Paging Quantization)是以 llama.cpp 為基礎的 CPU 推論路徑,讓 22.1 GB 的 Qwen3.6-35B-A3B-UD-Q4KM.gguf 在路由專家權重無法全部放入 RAM 時仍可執行。 GGUF 權重保留在儲存裝置,只有當前選中的專家會分頁載入容量受限的 RAM arena, MoE 專家計算則由 CPU 執行。本 repo 收錄程式碼與量測結果,不包含模型權重。 中文描述名:Qwen3-35B-A3B 專家分頁 4-bit 量化 CPU 推理版。…

Parameters—
Context—
Weights841.8 KB
Licenseapache-2.0
AccessOpen weights
Monthly Downloads—

Model Card

By Max Sun, published under apache-2.0, revision 8f2c3f5570c0.

EPQ(Expert Paging Quantization)是以 llama.cpp 為基礎的 CPU 推論路徑,讓 22.1 GB 的 Qwen3.6-35B-A3B-UD-Q4KM.gguf 在路由專家權重無法全部放入 RAM 時仍可執行。 GGUF 權重保留在儲存裝置,只有當前選中的專家會分頁載入容量受限的 RAM arena, MoE 專家計算則由 CPU 執行。本 repo 收錄程式碼與量測結果,不包含模型權重。 中文描述名:Qwen3-35B-A3B 專家分頁 4-bit 量化 CPU 推理版。 EPQ 的名稱刻意描述實際機制:MoE expert paging、一般 GGUF 權重量化,以及 CPU 低記憶體推理。本專案不是論文中的 SDQ(Sparse Decomposed Quantization): 沒有 2:4 結構化稀疏硬體跳零計算,也不宣稱使用 GPU 稀疏加速。sdq/ Python package、sdq C++ 檔案與 SDQ 除錯環境變數是歷史 implementation identifiers, 仍保留以避免破壞既有驗證工具;對外產品名稱與執行入口已改為 EPQ。 需求:Linux x86-64、C++ 工具鏈、CMake、Git、Python 3、足以存放 22.1 GB GGUF 與建置檔的磁碟空間,以及可提供實用解碼速度的本機高速 SSD/NVMe。已測試路徑使用 llama.cpp commit 836d57176dc699a726c55418e4f96b8ca628e1bf。 在其他位置執行時,此腳本會取得本 repo…

Read Max Sun's full model card

EPQ:Qwen3.6-35B-A3B 專家分頁與量化 CPU 推理

EPQ(Expert Paging Quantization)是以 llama.cpp 為基礎的 CPU 推論路徑,讓 22.1 GB 的 Qwen3.6-35B-A3B-UD-Q4_K_M.gguf 在路由專家權重無法全部放入 RAM 時仍可執行。 GGUF 權重保留在儲存裝置,只有當前選中的專家會分頁載入容量受限的 RAM arena, MoE 專家計算則由 CPU 執行。本 repo 收錄程式碼與量測結果,不包含模型權重。

中文描述名:Qwen3-35B-A3B 專家分頁 4-bit 量化 CPU 推理版。

EPQ 的名稱刻意描述實際機制:MoE expert paging、一般 GGUF 權重量化,以及 CPU 低記憶體推理。本專案不是論文中的 SDQ(Sparse Decomposed Quantization): 沒有 2:4 結構化稀疏硬體跳零計算,也不宣稱使用 GPU 稀疏加速。sdq/ Python package、sdq_* C++ 檔案與 SDQ_* 除錯環境變數是歷史 implementation identifiers, 仍保留以避免破壞既有驗證工具;對外產品名稱與執行入口已改為 EPQ。

快速開始

需求:Linux x86-64、C++ 工具鏈、CMake、Git、Python 3、足以存放 22.1 GB GGUF 與建置檔的磁碟空間,以及可提供實用解碼速度的本機高速 SSD/NVMe。已測試路徑使用 llama.cpp commit 836d57176dc699a726c55418e4f96b8ca628e1bf。

python3 demo.py --once

在其他位置執行時,此腳本會取得本 repo 原始碼,從 unsloth/Qwen3.6-35B-A3B-GGUF 下載 GGUF,建置指定版本的 llama.cpp,然後執行 短篇 CPU 推論。若模型已下載,可使用:

python3 demo.py --model-path /path/to/Qwen3.6-35B-A3B-UD-Q4_K_M.gguf \
  --once --ctx 4096 --threads 8 --n-predict 48

互動式聊天:

python3 demo.py --model-path /path/to/Qwen3.6-35B-A3B-UD-Q4_K_M.gguf \
  --ctx 4096 --threads 8

此 GGUF 不含 vision projector;本 EPQ 路徑僅支援文字推論。已量測設定為 4K context,在此 RAM 預算下尚未驗證模型宣稱的更長 context。

記憶體配置方式

模型 metadata 顯示每個 MoE 層有 256 個路由專家,每個 token 選用 8 個專家。GGUF 約 22.1 GB(十進位),其中路由專家權重約 18.2 GiB。EPQ 將未啟用的專家留在檔案 中,按需求把選中的專家載入記憶體 slab cache。此實作使用量化 GGUF 專家權重與 ggml CPU 向量點積 kernel、需求觸發的 LRU 式替換、平行讀取各專家不連續的 gate/up/down 權重區段,以及可與 CPU 計算重疊的專家計算 pipeline。

量測所用建置設定為 ubatch=128、4K context、greedy sampling;測試使用 8 條 CPU threads。依專家權重大小配置 slab,避免少數較大的專家浪費 arena 空間。跨層預取 預設關閉:在已測 workload 中,它增加流量並使速度退步。此 GGUF 沒有 MTP head; 在已測散文提示中,prompt-lookup n-gram drafting 的接受數為 0。

與傳統 llama.cpp 的同一 GGUF 比較

EPQ 並不是另一個模型。以下比較使用完全相同的 Qwen3.6-35B-A3B-UD-Q4_K_M.gguf,對照未修改的 llama.cpp CPU 路徑。

在同一台主機、8 threads、4K context、3 輪 decode throughput 中位數測試中:

執行路徑 Decode 中位數 相對結果
傳統 llama.cpp llama-completion 12.27 tok/s 1.00x
EPQ epq-chat 6.32 tok/s 0.515x

因此,在這組測試條件下,EPQ 比傳統 llama.cpp 慢約 48.5%;換一種說法, 傳統 llama.cpp 約為 EPQ 的 1.94 倍 decode 速度。

這個速度差是以 EPQ 將專家權重限制在約 8 GiB pager budget 的部署目標換來的: 傳統路徑讓整個 MoE 權重透過 mmap/CPU 路徑使用,而 EPQ 將選中的專家從 GGUF 按需讀入約 3.68 GiB arena。因此不能把結果解讀成 EPQ 的純計算 kernel 比 llama.cpp 慢 48.5%。

測試條件

  • 測試日期:2026-10-09
  • 主機:2× Intel Xeon Platinum 8559C,192 logical CPUs,AVX-512/AMX
  • 建置:Linux x86-64,Release,GGML_NATIVE=ON
  • llama.cpp commit:836d57176dc699a726c55418e4f96b8ca628e1bf
  • 模型:Qwen3.6-35B-A3B-UD-Q4_K_M.gguf,22,134,528,992 bytes
  • CPU threads:8;batch:4096;ubatch:128;context:4096
  • 生成:48 tokens,temperature 0,seed 42,純 completion,無 GPU
  • prompt:Write a short paragraph about how mixture-of-experts models route tokens.
  • 每個路徑執行 3 輪,表格採 decode throughput 中位數
  • 每輪獨立啟動程序;模型檔相同,未把 startup/model-load time 算入 decode tok/s

傳統 llama.cpp 指令(未套 EPQ patch):

llama-completion \
  -m Qwen3.6-35B-A3B-UD-Q4_K_M.gguf \
  -p "Write a short paragraph about how mixture-of-experts models route tokens." \
  -n 48 -c 4096 -t 8 -tb 8 -b 4096 -ub 128 \
  --temp 0 --seed 42 --no-warmup --no-display-prompt --no-jinja -no-cnv

EPQ 指令:

epq-chat \
  -m Qwen3.6-35B-A3B-UD-Q4_K_M.gguf \
  -p "Write a short paragraph about how mixture-of-experts models route tokens." \
  -n 48 -c 4096 -t 8 -b 128 \
  --temp 0 --no-chat --no-interactive --ram-budget-mb 8192

原始輪次

路徑 Run 1 Run 2 Run 3 中位數
傳統 llama.cpp 12.27 9.98 12.60 12.27 tok/s
EPQ 6.45 6.32 5.78 6.32 tok/s

llama.cpp 的統計以 47 個 decode runs 計算,因為第一個生成 token 不在該計時段; EPQ 同樣報告 47 decode tokens。EPQ CLI 顯示 15 prompt tokens,傳統 llama.cpp tokenizer 在相同字串下顯示 14 tokens;這只影響 prefill,不影響上述 decode throughput 比較。

EPQ 額外指標

指標 結果
Expert-cache hit rate 63.3%
Expert SSD read 約 7,813 MB / run,約 166.2 MB/decode token
Expert arena 約 3.68 GiB
Peak/ending RSS 約 7,618 MiB
Process VmSwap 0 MB(該行程觀測值)

VmSwap=0 MB 不代表主機停用 swap;本次主機有系統 swap,也不是嚴格的 8 GiB cgroup/no-swap 驗證。EPQ 的 --ram-budget-mb 8192 是 pager sizing budget, 不是整個程序的 RSS 硬上限。要驗證部署保證,仍需使用 memory.max、 memory.swap.max=0 的獨立環境。

完整的中文速度比較報告見 docs/llama_cpp_speed_comparison_zh.md。

不同 RAM 預算比較:4 / 8 / 16 / 32 / 60 GiB

在相同 prompt、8 threads、ctx 4096、ubatch 128、48-token greedy decode、每個 設定 3 輪的測試中,EPQ 的最佳結果是 16 GiB pager budget:6.70 tok/s。 RAM 越大不一定持續變快:16 GiB 後工作集已大部分常駐,32/60 GiB 的差異主要受 CPU、SSD 與單輪執行波動影響。

RAM pager budget 實際 arena Decode 中位數 相對 8 GiB Expert cache hit SSD 讀取 / token 結束 RSS
4 GiB 0.50 GiB 3.38 tok/s 0.62x 0.0% 599.4 MB 4,370 MiB
8 GiB 3.68 GiB 5.47 tok/s 1.00x 63.3% 166.2 MB 7,616 MiB
16 GiB 11.68 GiB 6.70 tok/s 1.23x 74.9% 86.1 MB 15,819 MiB
32 GiB 27.68 GiB 6.27 tok/s 1.15x 74.2% 91.3 MB 32,192 MiB
60 GiB 55.68 GiB 5.93 tok/s 1.08x 74.9% 86.1 MB 60,877 MiB

原始三輪 decode 結果:

RAM pager budget Run 1 Run 2 Run 3 中位數
4 GiB 3.39 3.20 3.38 3.38 tok/s
8 GiB 6.15 5.12 5.47 5.47 tok/s
16 GiB 6.99 6.70 5.38 6.70 tok/s
32 GiB 6.35 6.27 5.65 6.27 tok/s
60 GiB 5.93 5.22 6.32 5.93 tok/s

以同機傳統 llama.cpp 的 12.27 tok/s 為基準,4/8/16/32/60 GiB 分別約為 基準的 27.5%、44.6%、54.6%、51.1%、48.4%。這些是單一主機、單一 prompt、 3 輪的工程測試,不是跨硬體保證。

注意:--ram-budget-mb 是 EPQ pager sizing budget,不是 OS/cgroup 的總 RSS 硬限制。60 GiB 設定本次實測結束 RSS 約 60.9 GB,請確保主機有足夠實體記憶體, 並且不要把這些測試解讀為 no-swap 驗證。完整測試條件與逐輪結果見上述中文報告。

歷史速度與記憶體結果

以下是 repo 中既有的歷史量測,不是上面同一 GGUF 對照測試的重複結果:

量測項目 結果 條件
Decode throughput 中位數 6.28 tokens/s 7 runs、8 threads、ctx 4096、ubatch 128、48 generated tokens
Prefill throughput 約 30 tokens/s 681-token prompt、ubatch 128
Expert-cache hit rate 70.7% representative decode workload
Expert I/O 約 202 MB/token 十進位 MB,視 route/cache state 而異
Peak process RSS 8,035 MiB 長 prompt/4K-context 壓力測試
Process VmSwap 0 MB 該幾輪量測的觀察值

以上數值來自 docs/03_results.md。不同階段的 benchmark 主機與方法不同:較早的 速度測試使用 8 GiB RLIMIT_DATA 上限,後續測試則是在較寬鬆的外層限制下將 RAM 預算交給 pager。RLIMIT_DATA 限制 data/heap 配置,不是總常駐記憶體;它不會 停用系統 swap,也不能證明總 RSS 低於 8 GiB。請勿將上述數值解讀為嚴格 8 GiB 實體 RAM/no-swap 驗證。SSD 速度、記憶體壓力、CPU quota、檔案系統快取與 prompt 都會影響 tok/s。

設計依據與實測取捨

此設計利用 MoE 的稀疏專家啟用特性:路由專家會動態選取,因此計算 token 時不需要 把全部權重常駐 RAM。DeepSeekMoE 討論稀疏專家啟用與 shared experts (Dai 等人,2024);PagedAttention 將分頁用於 serving 系統的 KV cache 管理(Kwon 等人,SOSP 2023)。 這些研究提供記憶體階層設計的背景,但都沒有評估此 CPU/GGUF 實作。

技術 本地證據/決策
將專家快取限制在 RAM;未啟用專家留在檔案 核心設計;另一組 2-thread 測試中,8 GiB pager 預算結果為 3.45 tok/s。
平行讀取 gate/up/down 既有 5 輪比較:5.52 vs 3.73 tok/s,約提升 48%;結果依主機與儲存裝置而異。
依權重大小配置槽位 既有 7 輪比較:6.28 vs 5.63 tok/s,約提升 11.5%,RSS 相近。
Speculative decoding 此 GGUF 沒有 MTP 權重;本地 n-gram replay 在已測散文提示中的接受數為 0,因此未啟用。
跨層專家預取 本地結果速度退步 8% 且 I/O 增加,故預設關閉。
對已量化熱專家做無損壓縮 本地 Zstd/LZ4 測試僅回收約 0–2.5% 空間,未用於執行期快取。
更積極的有損專家替代/剪枝 可能以品質換吞吐量;此處尚未驗證輸出品質,因此未啟用。

最佳化目標是在不更動路由或權重的前提下縮短每個 token 的關鍵路徑。若實驗改變 取樣輸出,必須另行評估品質,才能視為速度提升。

Benchmark 與驗證指令

# 基準測試(假設模型已下載)
python3 -m sdq.bench --model /path/to/model.gguf \
  --bin /path/to/llama.cpp/build/bin/epq-chat --ram-mb 8192 --rounds 2

# 比較儲存裝置並行度與槽位大小策略
python3 -m sdq.experiments --model /path/to/model.gguf \
  --bin /path/to/llama.cpp/build/bin/epq-chat --suite io --repeat 3
python3 -m sdq.experiments --model /path/to/model.gguf \
  --bin /path/to/llama.cpp/build/bin/epq-chat --suite slot --repeat 7

# 檢查目標儲存裝置;direct I/O 支援與隨機讀取表現很重要
python3 tools/io_probe.py --model /path/to/model.gguf

# kernel/layer 檢查與 logits 確定性檢查
python3 demo.py --model-path /path/to/model.gguf --verify
python3 tools/determinism_check.py --model /path/to/model.gguf \
  --bin /path/to/llama.cpp/build/bin/epq-chat --runs 5 -n 48

sdq.bench 是 legacy Python package 名稱,預設會套用 RLIMIT_DATA;這不是硬性 RSS 或 no-swap 限制。不要將它 解讀為 OS 記憶體上限。請同時回報 VmRSS/peak RSS 與 VmSwap,若要驗證嚴格部署 限制,請使用停用 swap 的 OS-level cgroup。

Repository layout

  • cpp/、patches/:C++ expert pager 與 llama.cpp 整合。
  • sdq/:設定、benchmark、實驗、驗證與 HF sync。
  • tools/:GGUF 檢查、正確性檢查、route/storage probe。
  • docs/:設計、完整量測結果、負面結果與環境稽核。
  • stage.txt:依時間排列的專案狀態。

License

Qwen base model 由 upstream 以 Apache-2.0 發布;請以 upstream model card 的目前 條款為準。本 repository 包含推論軟體與量測結果,不包含模型權重。llama.cpp 與其 相依套件保留各自授權條款。

Identity and Version

Repository
HelloSun/qwen35a3b-moe-expert-paging-q4
Publisher
Max Sun
Task
Text generation
Modality
Text
Library
gguf
Parameters
Not stated by the source
Languages
moe, cpu, low-memory
Revision
8f2c3f5570c0a4e58d2ad5e9ac4e0790534705a5
First published
2026-10-03
Last updated
2026-10-10

Files and Weights

85 files, 841.8 KB in total.

Configuration32 files · 184.5 KB
Documentation28 files · 118.5 KB
Other23 files · 537.2 KB
Repository2 files · 1.6 KB
Every file
FileTypeSizeSHA-256
bench/expert_index.jsonConfiguration33.2 KB —
bench/results/bench_first.jsonConfiguration6.8 KB —
bench/results/bench_pipeline.jsonConfiguration6.8 KB —
bench/results/bench_s2.jsonConfiguration6.5 KB —
bench/results/bench_s4_modec.jsonConfiguration6.4 KB —
bench/results/exp_final_ab.jsonConfiguration2.7 KB —
bench/results/exp_rest_modec.jsonConfiguration4.1 KB —
bench/results/exp_s5_io.jsonConfiguration3.1 KB —
bench/results/exp_s5_io2.jsonConfiguration4.0 KB —
bench/results/exp_s6_slot_ab.jsonConfiguration1.1 KB —
bench/results/exp_s7_pf.jsonConfiguration2.1 KB —
bench/results/exp_s7_pf_confirm.jsonConfiguration1.1 KB —
bench/results/exp_s8_release.jsonConfiguration1.1 KB —
bench/results/exp_s9_final.jsonConfiguration1.1 KB —
bench/results/exp_s9_final2.jsonConfiguration1.1 KB —
bench/results/exp_s9_final3.jsonConfiguration1.3 KB —
bench/results/exp_s9_racefix.jsonConfiguration1.2 KB —
bench/results/exp_smoke.jsonConfiguration884 B —
demo.pyConfiguration15.2 KB —
sdq/__init__.pyConfiguration —
sdq/bench.pyConfiguration6.6 KB —
sdq/config.pyConfiguration2.4 KB —
sdq/experiments.pyConfiguration17.1 KB —
sdq/hf_sync.pyConfiguration7.3 KB —
sdq/verify.pyConfiguration12.8 KB —
tools/determinism_check.pyConfiguration5.3 KB —
tools/drop_model_cache.pyConfiguration1.6 KB —
tools/gguf_index.pyConfiguration10.1 KB —
tools/io_probe.pyConfiguration4.3 KB —
tools/moe_layer_check.pyConfiguration3.0 KB —
tools/moe_ref.pyConfiguration9.2 KB —
tools/route_overlap.pyConfiguration5.2 KB —
README.mdDocumentation13.0 KB —
bench/results/exp_final_ab.mdDocumentation715 B —
bench/results/exp_rest_modec.mdDocumentation939 B —
bench/results/exp_s5_io.mdDocumentation875 B —
bench/results/exp_s5_io2.mdDocumentation1.1 KB —
bench/results/exp_s6_slot_ab.mdDocumentation398 B —
bench/results/exp_s7_pf.mdDocumentation584 B —
bench/results/exp_s7_pf_confirm.mdDocumentation383 B —
bench/results/exp_s8_release.mdDocumentation405 B —
bench/results/exp_s9_final.mdDocumentation405 B —
bench/results/exp_s9_final2.mdDocumentation424 B —
bench/results/exp_s9_final3.mdDocumentation405 B —
bench/results/exp_s9_racefix.mdDocumentation393 B —
bench/results/exp_smoke.mdDocumentation382 B —
docs/01_moe_path_analysis.mdDocumentation5.9 KB —
docs/02_design.mdDocumentation6.7 KB —
docs/03_results.mdDocumentation6.3 KB —
docs/04_experiments.mdDocumentation13.9 KB —
docs/05_optimization.mdDocumentation11.0 KB —
docs/06_slots.mdDocumentation7.2 KB —
docs/07_prefetch.mdDocumentation6.4 KB —
docs/08_verify_gaps.mdDocumentation6.4 KB —
docs/08gb_no_swap.mdDocumentation10.9 KB —
docs/09_determinism.mdDocumentation7.2 KB —
docs/10_activation_cache_negative.mdDocumentation3.9 KB —
docs/11_final_barrier_negative.mdDocumentation2.8 KB —
docs/12_tiled_adapter_negative.mdDocumentation3.2 KB —
docs/llama_cpp_speed_comparison_zh.mdDocumentation6.3 KB —
bench/results/exp_s5_io_final.txtOther463 B —
bench/results/exp_s5_io_run1.logOther1.8 KB —
bench/results/exp_s6_slot_ab.logOther971 B —
bench/results/exp_s7_pf.logOther1.3 KB —
bench/results/route_decode48.logOther89.5 KB —
bench/results/s9_det2.logOther418 B —
bench/results/s9_final3.logOther1.0 KB —
bench/results/s9_race.logOther1.4 KB —
bench/results/s9_racefix.logOther2.4 KB —
bench/results/stats_cold.jsonlOther1.0 KB —
bench/results/stats_hot0.jsonlOther1.0 KB —
bench/results/stats_hot1.jsonlOther1.0 KB —
bench/results/verify_s6.logOther751 B —
bench/results/verify_s7.logOther1.7 KB —
bench/results/verify_s8_full.logOther2.2 KB —
bench/results/verify_s8_pfstress.logOther461 B —
cpp/sdq_buft.cppOther8.0 KB —
cpp/sdq_cli.cppOther31.9 KB —
cpp/sdq_moe.cppOther32.7 KB —
cpp/sdq_pager.cppOther74.6 KB —
cpp/sdq_pager.hOther19.1 KB —
patches/0001-sdq-expert-pager.patchOther176.5 KB —
stage.txtOther86.9 KB —
.gitattributesRepository1.5 KB —
.gitignoreRepository72 B —

License and Download

License
apache-2.0
Access
Open weights, no gate
Download from Max Sun

Released by Max Sun through its official repository on Hugging Face. Read the license.

Built From

Questions About qwen35a3b-moe-expert-paging-q4

Can I use qwen35a3b-moe-expert-paging-q4 commercially?

Yes. qwen35a3b-moe-expert-paging-q4 is released under Apache License 2.0. The Apache License 2.0 is a permissive open-source license. It permits commercial use, modification and redistribution. It requires keeping the license and copyright notices and any NOTICE file, stating significant changes, and it includes an express patent grant from contributors.

Similar Models

Model · Text generation

opt-125m

AI at Meta

OPT was first introduced in Open Pre-trained Transformer Language Models and first released in metaseq's repository on May 3rd 2022 by Meta AI. Disclaimer: The team releasing OPT wrote an official model card, which is available in Appendix D of the paper. Content from this model card has been written by the Hugging Face team. To quote the first two paragraphs of the official paper OPT was predominantly pretrained with English text, but a small amount of non-English data is still present within the training corpus via CommonCrawl. The model was pretrained using a causal language modeling (CLM) objective. OPT belongs to the same family of decoder-only models like GPT-3. As such, it was…

Open weights other 2,048 tokens transformers

Model · Text generation

Ornith-1.5-9B-GGUF

Ornith

Chirp Chirp! We are introducing Ornith-1.5, a major step toward building foundation models through end-to-end self-improvement. Ornith-1.5 extends Ornith-1.0 (which was developed on top of Qwen3.5 and Gemma4 with additional continued pretraining, mid-training, and post-training) by expanding the self-improvement loop from scaffold and rollout optimization to jointly optimizing task generation, scaffold construction, and solution rollouts. Rather than relying on a fixed set of human-curated tasks and manually designed harnesses, Ornith-1.5 continuously generates new training tasks, discovers effective strategies for solving them, and improves the policy through reinforcement learning. For…

Open weights mit transformers

Model · Text generation

Ornith-1.5-35B-A3B-GGUF

Ornith

Chirp Chirp! We are introducing Ornith-1.5, a major step toward building foundation models through end-to-end self-improvement. Ornith-1.5 extends Ornith-1.0 (which was developed on top of Qwen3.5 and Gemma4 with additional continued pretraining, mid-training, and post-training) by expanding the self-improvement loop from scaffold and rollout optimization to jointly optimizing task generation, scaffold construction, and solution rollouts. Rather than relying on a fixed set of human-curated tasks and manually designed harnesses, Ornith-1.5 continuously generates new training tasks, discovers effective strategies for solving them, and improves the policy through reinforcement learning. For…

Open weights mit transformers

Model · Text generation

Ternary-Bonsai-2-27B-gguf

Prism ML

Full 27B-class reasoning in ternary transformer weights, for llama.cpp (CUDA, Metal, CPU) - \~5.9 GB language model (down from \~54 GB FP16) — full 27B-class reasoning on a standard laptop or a single GPU - 98.2% of FP16 intelligence retained: 84.78 average across 14 thinking-mode benchmarks — far above the conventional IQ2XXS build (72.59) at about 82% of its footprint, and within 0.4 points of UD-Q4KXL at three times the footprint - Retains thinking, reasoning, and agentic behavior deep in the sub-4-bit regime, where conventional low-bit representations collapse: math within half a point of full precision (96.57), coding level with the baseline (89.42), agentic tool calling at 74.92…

Open weights apache-2.0 llama.cpp

Uncensored Qwen3.8-27B, published as GGUF quantizations with the multi token prediction (MTP) head retained and verified. Refusal behaviour has been substantially reduced, not eliminated. See Measured behaviour for the numbers. Capabilities, training data, and architecture are otherwise unchanged. - Refusal directions removed with Heretic, which co minimizes refusal count against KL divergence from the base model. No handwritten refusal removal code, no finetuning, no additional training data. - Abliteration runs at bf16 (no 4 bit quantization). the resulting LoRA is merged into the bf16 base, so the published weights are not a quantized round trip. - mtp. tensors are copied verbatim from…

Open weights apache-2.0 llama.cpp

Model · Text generation

Ornith-1.0-9B-GGUF

Ornith

Aloha! Today, we are releasing Ornith-1.0, a self-improving family of open-source models for agentic coding. This model card documents Ornith-1.0-9B, the most lightweight member of the Ornith family, designed for efficient single-GPU deployment. Ornith-1.0-9B is a dense ~9B model (≈19 GB in bf16), so it serves comfortably on a single 80GB GPU. The recipes below stand up an OpenAI-compatible server; add --tensor-parallel-size / --tp if you want to shard across more GPUs. For a quick local test (or to script offline generation), load the model directly with Transformers. Make sure you have a recent release installed — see the Transformers installation guide; Ornith-1.0-9B requires…

Open weights mit transformers