Open-Weight Models

權重與大 KV cache 放在 SSD 檔案，RAM 只留執行必須的部分。禁止 swap。 CPU 純推論（無 GPU）。 - 模型：unsloth/Qwen3.5-4B-GGUF / Qwen3.5-4B-Q4KM.gguf（2.553 GiB） - 引擎：llama.cpp commit 0504396 + patches/0001-ssd-mmap-release.patch + patches/0002-kv-ssd-pager.patch - 硬約束：行程總 RSS（含 page cache）≤ 1 GiB；VmSwap = 0

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（pinned commit，見 sdq/config.py） - 核心實作：cpp/sdqpager.cpp（RAM arena + 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） - 文件：docs/01 執行路徑分析、docs/02 設計、docs/03 實測結果與瓶頸、 docs/04 mode C 與負結果、docs/05 I/O 段平行與 8 GB 預算修正 - 一鍵從零到聊天：python demo.py 條件：--ram-budget-mb 8192 + RLIMITDATA 8 GB、8 執行緒、ctx 4096、temp 0， 短提示詞、48 個生成 token。換機器數字一定不同（見 docs/05 §1：本機的 SSD 上下文設計上支援 1M，實測只驗 4K。細節見 docs/03results.md（量測方法）、 docs/04experiments.md（mode C 與兩個負結果）、docs/05optimization.md（本輪的 一句話：一個 expert 的三段權重（gate/up/down）在…

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。…

Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference

Two Qwen3-4B full-finetunes that judge one proposed agent action, before it executes, against the observed prefix and the task policy, and emit a single verdict box. They differ only in SFT learning rate. act is CONTINUE / ASK / STOP. The binary projection is CONTINUE=SAFE, STOP=UNSAFE, and ASK=UNSAFE at risk >= 50. cite points at an earlier visible step, or NONE. fail / harm / src come from a 16 failure-mode, 11 harm-type, 10 risk-source taxonomy; NONE is an absence sentinel, not an additional class. Full SFT from Qwen/Qwen3-4B, epoch 2, thinking enabled, on 2,464 step-level records. Prompts were rendered through the evaluator's own pipeline, so the training and inference formats match.…

Model

sdxl-unet-q8-gguf

HYBRIDION

SDXL UNet only, in four self-describing GGUF variants. Text encoders and VAE are not in these files — they come from any SDXL checkpoint. All four carry the same self-description KVs and load through the same patch/loader. CPU and CPU-offload deployments should pick Q40 (fastest download, same runtime speed as the others once decoded). The file carries its own integration metadata as GGUF string KVs: Loaders read only the header to classify the file — no weight scan. Requires the SDXL-GGUF patch (config + loader + CUDA dequant dispatch): https://github.com/aeuphoraex/invokeai-sdxl-gguf The loader fetches text encoders + VAE from the embedded source-checkpoint path, falling back to any…

A bundle of 12 frozen Graph Transformer Network checkpoints, four per training seed 11, 22, 33, trained on SEAL 5-component D9 data: - gamma: physical voltage-conversion ratio regressor. - efficiency: physical efficiency regressor; together with gamma, this is the frozen Loss-Net. - mse: MSE-only duty Task-Net. - seal: SEAL-Static duty Task-Net, trained with frozen physical-model Utility guidance. The original checkpoint filenames and selected model hashes come from the completed E1 model locks. No retraining, averaging, or Test/MCTS reselection was performed for publication. D5 and GBI are duty-selection procedures using the same physical pair; they do not have separate trained…

GGUF build of ironbcc/seb-9b for llama.cpp, Ollama and other GGUF runtimes. Seb-9B is a local multimodal model for fast decisions: classification, routing and scoring, with probabilities over your options. Accuracy, calibration and limitations are on the main model card. Seb answers one decision per request. Send the question as compact JSON in the user message and read back one key: - type: noul (keys Y/N), choice (keys A..T, up to 20) or score (keys 0..4). - state: the data to decide on. For an image, attach it to the message and put "image":"attached" in state. - criteria (optional): rules the decision must follow. - Turn thinking off, generate one token at temperature 0, and request…

Model

sebot

Precontation

i thought it would be funny to train an ai off of every single message by a random guy on discord and... yeah it was worth it thanks to Sebel"son"zejdhfe (aka sebelzahnziege) for sort of letting me steal his likeness - Qwen3-8B.Q4KM.gguf

This museglimmer model was trained 2x faster with Unsloth

weighted/imatrix quants of https://huggingface.co/Taimwe/securecoder-30b-pro-merged For a convenient overview and download list, visit our model page for this model. static quants are available at https://huggingface.co/mradermacher/securecoder-30b-pro-merged-GGUF If you are unsure how to use GGUF files, refer to one of TheBloke's READMEs for more details, including on how to concatenate multi-part files. (sorted by size, not necessarily quality. IQ-quants are often preferable over similar sized non-IQ quants) Here is a handy graph by ikawrakow comparing some lower-quality quant And here are Artefact2's thoughts on the matter…

weighted/imatrix quants of https://huggingface.co/Taimwe/securecoder-30b-pro-v2-merged For a convenient overview and download list, visit our model page for this model. static quants are available at https://huggingface.co/mradermacher/securecoder-30b-pro-v2-merged-GGUF If you are unsure how to use GGUF files, refer to one of TheBloke's READMEs for more details, including on how to concatenate multi-part files. (sorted by size, not necessarily quality. IQ-quants are often preferable over similar sized non-IQ quants) Here is a handy graph by ikawrakow comparing some lower-quality quant And here are Artefact2's thoughts on the matter…

Model forensik Hunter berbasis ConvNeXtV2-Tiny yang dilatih murni dari awal (from scratch / zero pretraining) dengan mekanisme penalti asimetris untuk meminimalkan False Negative pada citra sintetis AI. - Weight Real (0): 1.0 - Weight AI (1): 3.0 (Hukuman 3x lipat jika AI lolos)

This lane contains 3 distinct original checkpoint variants. Each checkpoint uses a content-addressed path and an asset manifest preserving its schema, decoder role, available format/profile binding, codec, source declarations, and companion corpus or embedding assets. Unknown bindings remain unknown. Registration and publication do not qualify numerical quality, source-only reconstruction, teacher use, proof authority, or serving. No retained checkpoint is activated for inference. The Legal 768D checkpoint is inherited initialization only; Codebase, Intent, and Security 768D lanes have no retained checkpoints. Codebase 384D envelopes retain their borrowed Security payload identity and are…