# Models, Page 411 of 477 Source: https://savrn.com/models/page/411 Markdown alternate of the page above; the site index is https://savrn.com/llms.txt --- SAVRN Model Hub # Open-Weight Models An open-weight model is an AI model whose trained weights are published for anyone to download. The weights are what the model learned in training. With a copy of them you can run the model on hardware you control and train it further on your own data. Open weights are not the same as open source. Many publishers release the weights without the training data or code, and the license sets what you may do with the model. This library puts each model's full card, architecture, files, license and published evaluations on one page. 11,442Models 3,247Datasets 285Papers 6,976Publishers 12,091Sourced relationships Updated 2026-10-19 · [How the library is built](https://savrn.com/model-hub/methodology) ## Filter Models Task - [Text generation1,413](https://savrn.com/models?task=Text+generation) - [Image and text to text512](https://savrn.com/models?task=Image+and+text+to+text) - [Feature extraction341](https://savrn.com/models?task=Feature+extraction) - [Robotics338](https://savrn.com/models?task=Robotics) - [Text classification303](https://savrn.com/models?task=Text+classification) - [Text to image209](https://savrn.com/models?task=Text+to+image) - [Speech recognition171](https://savrn.com/models?task=Speech+recognition) - [Audio classification157](https://savrn.com/models?task=Audio+classification) - [Image classification137](https://savrn.com/models?task=Image+classification) - [Text to speech132](https://savrn.com/models?task=Text+to+speech) - [Sentence similarity103](https://savrn.com/models?task=Sentence+similarity) - [Token classification97](https://savrn.com/models?task=Token+classification) - [Reinforcement learning93](https://savrn.com/models?task=Reinforcement+learning) - [Zero-shot classification93](https://savrn.com/models?task=Zero-shot+classification) - [Image segmentation88](https://savrn.com/models?task=Image+segmentation) - [Object detection86](https://savrn.com/models?task=Object+detection) - [Any to any79](https://savrn.com/models?task=Any+to+any) - [Translation78](https://savrn.com/models?task=Translation) - [Question answering73](https://savrn.com/models?task=Question+answering) - [Fill mask69](https://savrn.com/models?task=Fill+mask) - [Time series forecasting64](https://savrn.com/models?task=Time+series+forecasting) - [Video classification59](https://savrn.com/models?task=Video+classification) - [Image to text58](https://savrn.com/models?task=Image+to+text) - [Text to video55](https://savrn.com/models?task=Text+to+video) Library - [transformers2,777](https://savrn.com/models?library=transformers) - [peft345](https://savrn.com/models?library=peft) - [diffusers187](https://savrn.com/models?library=diffusers) - [pytorch178](https://savrn.com/models?library=pytorch) - [gguf167](https://savrn.com/models?library=gguf) - [lerobot139](https://savrn.com/models?library=lerobot) - [sentence-transformers139](https://savrn.com/models?library=sentence-transformers) - [mlx136](https://savrn.com/models?library=mlx) - [timm49](https://savrn.com/models?library=timm) - [transformers.js45](https://savrn.com/models?library=transformers.js) - [flaxspeaker44](https://savrn.com/models?library=flaxspeaker) - [llama.cpp44](https://savrn.com/models?library=llama.cpp) - [vllm38](https://savrn.com/models?library=vllm) - [onnx36](https://savrn.com/models?library=onnx) - [diffusion-single-file27](https://savrn.com/models?library=diffusion-single-file) - [minimax-h324](https://savrn.com/models?library=minimax-h3) - [ultralytics24](https://savrn.com/models?library=ultralytics) - [PaddleOCR22](https://savrn.com/models?library=PaddleOCR) - [openvino18](https://savrn.com/models?library=openvino) - [coreml17](https://savrn.com/models?library=coreml) - [exllamav316](https://savrn.com/models?library=exllamav3) - [openpi16](https://savrn.com/models?library=openpi) - [litert15](https://savrn.com/models?library=litert) - [rkllm15](https://savrn.com/models?library=rkllm) Modality - [Other6,473](https://savrn.com/models?modality=Other) - [Text2,612](https://savrn.com/models?modality=Text) - [Image and text570](https://savrn.com/models?modality=Image+and+text) - [Image565](https://savrn.com/models?modality=Image) - [Audio478](https://savrn.com/models?modality=Audio) - [Control431](https://savrn.com/models?modality=Control) - [Video114](https://savrn.com/models?modality=Video) - [Multimodal79](https://savrn.com/models?modality=Multimodal) - [Time series64](https://savrn.com/models?modality=Time+series) - [Tabular56](https://savrn.com/models?modality=Tabular) License - [apache-2.03,105](https://savrn.com/models?license=apache-2.0) - [mit1,246](https://savrn.com/models?license=mit) - [other703](https://savrn.com/models?license=other) - [cc-by-nc-4.0193](https://savrn.com/models?license=cc-by-nc-4.0) - [cc-by-4.0148](https://savrn.com/models?license=cc-by-4.0) - [bsd-3-clause98](https://savrn.com/models?license=bsd-3-clause) - [gemma62](https://savrn.com/models?license=gemma) - [cc-by-nc-sa-4.049](https://savrn.com/models?license=cc-by-nc-sa-4.0) - [llama3.240](https://savrn.com/models?license=llama3.2) - [agpl-3.032](https://savrn.com/models?license=agpl-3.0) - [llama3.128](https://savrn.com/models?license=llama3.1) - [unknown27](https://savrn.com/models?license=unknown) - [cc-by-sa-4.020](https://savrn.com/models?license=cc-by-sa-4.0) - [gpl-3.017](https://savrn.com/models?license=gpl-3.0) - [openrail++17](https://savrn.com/models?license=openrail%2B%2B) - [openmdw-1.115](https://savrn.com/models?license=openmdw-1.1) - [llama314](https://savrn.com/models?license=llama3) - [openrail14](https://savrn.com/models?license=openrail) - [cc-by-nc-nd-4.013](https://savrn.com/models?license=cc-by-nc-nd-4.0) - [creativeml-openrail-m11](https://savrn.com/models?license=creativeml-openrail-m) - [afl-3.07](https://savrn.com/models?license=afl-3.0) - [llama27](https://savrn.com/models?license=llama2) - [osl-3.06](https://savrn.com/models?license=osl-3.0) - [cc0-1.04](https://savrn.com/models?license=cc0-1.0) Access - [Open weights11,087](https://savrn.com/models?access=open_access) - [Access requested at publisher355](https://savrn.com/models?access=gated) Publisher - [Team Mradermacher294](https://savrn.com/models?publisher=mradermacher) - [Yurun Yuan95](https://savrn.com/models?publisher=ryanyr) - [RunningHub.ai94](https://savrn.com/models?publisher=runninghubai) - [Qwen91](https://savrn.com/models?publisher=qwen) - [Jenny Wilson81](https://savrn.com/models?publisher=jenny08311) - [RKNNAI73](https://savrn.com/models?publisher=rknnai) - [AI at Meta72](https://savrn.com/models?publisher=facebook) - [Jessen Pu69](https://savrn.com/models?publisher=putty77) - [Convergent Intelligence66](https://savrn.com/models?publisher=reaperdoesntknow) - [Google64](https://savrn.com/models?publisher=google) - [Yurun Yuan63](https://savrn.com/models?publisher=yurunyyr) - [Firza Hadzami61](https://savrn.com/models?publisher=firzahdzm) - [Zeyuan Yang59](https://savrn.com/models?publisher=miiche) - [Yanggangu57](https://savrn.com/models?publisher=yanggangu) - [Hongao56](https://savrn.com/models?publisher=cisco1963) - [Francesca Padovani53](https://savrn.com/models?publisher=francesca9805) - [Wang49](https://savrn.com/models?publisher=davidwdw) - [Microsoft48](https://savrn.com/models?publisher=microsoft) - [Yurun Yuan48](https://savrn.com/models?publisher=yuruny) - [NVIDIA46](https://savrn.com/models?publisher=nvidia) - [Khazzz1c44](https://savrn.com/models?publisher=khazic) - [Quan44](https://savrn.com/models?publisher=wq2012) - [BAAR43](https://savrn.com/models?publisher=aixk) - [Christopher Subia-Waud42](https://savrn.com/models?publisher=cwaud) [Most downloaded](https://savrn.com/models?sort=downloads)[Most liked](https://savrn.com/models?sort=likes)[Recently updated](https://savrn.com/models?sort=updated)[Name](https://savrn.com/models?sort=name)[Largest](https://savrn.com/models?sort=params) 11,442 models, sorted by most downloaded. Model · Text generation ### [qwen3.8-flash-next-40b-prune-research](https://savrn.com/models/qwen3-8-flash-next-40b-prune-research) [DavidB](https://savrn.com/model-publishers/davd-b01) One upstream MoE, six generations of masks, one bit-exact surgery, a rebuilt factual memory, and a three-stage post-training (SFT → SimPO → RLVR) — with every error I made and every idea I falsified along the way. This repository is the complete technical record of the project: the activation-instrumentation artifacts, the mask-generation methodology, the surgery code, the PLE (n-gram memory) rebuild, the trainers, and the full lab notebook. The champion weights live in sibling repos What I set out to do: take Qwen3.8-Flash-Next — an ~180B-class MoE (~130B of core network + 51B of n-gram memory table + 4B MTP) — and cut it by 75% of its experts (512 → 128) while conserving the maximum… Open weights apache-2.0 40.7B parameters 262,144 tokens transformers [View model](https://savrn.com/models/qwen3-8-flash-next-40b-prune-research) Model · Text generation ### [qwen3.8-flash-next-radiance-q5-w5dn64](https://savrn.com/models/qwen3-8-flash-next-radiance-q5-w5dn64) [Gostar](https://savrn.com/model-publishers/gostar77) A checkpoint-specific calibrated quantization of Qwen/Qwen3.8-Flash-Next for the Radiance inference engine. This repository contains a single self-contained.rad inference container. It is not a Transformers, safetensors, llama.cpp, or vLLM checkpoint. Use a Radiance build with the Qwen3.8-Flash-Next model support and w4nl/w5nl expert kernels. - All 48 calibrated trunk layers use GPTQ with three coordinate-descent sweeps. - Routed expert down projections use non-linear w5nl (5-bit) codebooks, group 64, E4M3 scales, and FWHT128. - Routed expert gate/up projections use w4nl (4-bit) with the same transform and group structure. - Ten selected high-energy routed experts remain BF16. - Attention… Open weights other radiance [View model](https://savrn.com/models/qwen3-8-flash-next-radiance-q5-w5dn64) Model ### [qwen3.8-flash-next_p150x4](https://savrn.com/models/qwen3-8-flash-next-p150x4) [Sam Jett](https://savrn.com/model-publishers/sjetttt) Qwen3.8-Flash-Next (48 hybrid layers: gated delta-net, sparse attention with an indexer, a 512-expert top-10 MoE at BF4, plus a PLE n-gram layer) served through vLLM and the Tenstorrent vLLM plugin on 4x p150 (a QuietBox or four p150 cards) or on a QuietBox 2 (2x p300c). One interactive user at a time; contexts of 32k, 64k or 128k tokens; thinking off by default. Updated 2026-09-26: prefill in 2048-row slabs by default (1,990 prompt tokens/s), decode 29.1 tokens/s greedy / 27.4 sampled at batch 1 (fork commit ce5b0f8eb6eb). Runs on 4x p150 (a QuietBox or four p150 cards; profiles c32k, c64k, c128k, c32k-quietbox) and on a QuietBox 2 (2x p300c; the default c32k or c32k-quietbox2) — see the… Open weights [View model](https://savrn.com/models/qwen3-8-flash-next-p150x4) Model ### [qwen3.8_27b_TURBO_Fable_Cold_Fusion_uncensored_bf16.safetensors](https://savrn.com/models/qwen3-8-27b-turbo-fable-cold-fusion-uncensored-bf16-safetensors) [Maimaitiming](https://savrn.com/model-publishers/memtimen) Open weights [View model](https://savrn.com/models/qwen3-8-27b-turbo-fable-cold-fusion-uncensored-bf16-safetensors) Model · Image and text to text ### [qwen35-9b-harvey-v4-notes-ntp-100m](https://savrn.com/models/qwen35-9b-harvey-v4-notes-ntp-100m) [Violet Xiang](https://savrn.com/model-publishers/violetxi) Full Qwen3.5-9B model from the September 22, 2026 Harvey notes-only training runs. This revision is epoch 2, step 1550 of a two-epoch run, job 1016041. The final checkpoint is on main and final; epoch 1 is on epoch1. Each saved checkpoint is also available by its checkpoint- reference. - 100,000,076 note labels/epoch before causal shift; 99,993,882 scored labels/epoch after shift. - 223,681 notes, packed into 6,200 rows of 16,384 tokens with seed 731. - 199,987,764 scored supervised labels seen through this checkpoint. - No supervised trajectory examples. These are the notes-only controls, separate from notes + trajectory mixture models. - Eight GH200 GPUs, effective batch 8, learning rate… Open weights apache-2.0 9.7B parameters 262,144 tokens transformers [View model](https://savrn.com/models/qwen35-9b-harvey-v4-notes-ntp-100m) Model · Image and text to text ### [qwen35-9b-harvey-v4-notes-ntp-100m-kl-0p01](https://savrn.com/models/qwen35-9b-harvey-v4-notes-ntp-100m-kl-0p01) [Violet Xiang](https://savrn.com/model-publishers/violetxi) Full Qwen3.5-9B model from the September 22, 2026 Harvey notes-only training runs. This revision is epoch 2, step 1550 of a two-epoch run, job 1016164. The final checkpoint is on main and final; epoch 1 is on epoch1. Each saved checkpoint is also available by its checkpoint- reference. - 100,000,076 note labels/epoch before causal shift; 99,993,882 scored labels/epoch after shift. - 223,681 notes, packed into 6,200 rows of 16,384 tokens with seed 731. - 199,987,764 scored supervised labels seen through this checkpoint. - No supervised trajectory examples. These are the notes-only controls, separate from notes + trajectory mixture models. - Eight GH200 GPUs, effective batch 8, learning rate… Open weights apache-2.0 9.7B parameters 262,144 tokens transformers [View model](https://savrn.com/models/qwen35-9b-harvey-v4-notes-ntp-100m-kl-0p01) Model · Image and text to text ### [qwen35-9b-harvey-v4-notes-ntp-30m](https://savrn.com/models/qwen35-9b-harvey-v4-notes-ntp-30m) [Violet Xiang](https://savrn.com/model-publishers/violetxi) Full Qwen3.5-9B model from the September 22, 2026 Harvey notes-only training runs. This revision is epoch 2, step 466 of a two-epoch run, job 1016040. The final checkpoint is on main and final; epoch 1 is on epoch1. Each saved checkpoint is also available by its checkpoint- reference. - 29,999,869 note labels/epoch before causal shift; 29,998,010 scored labels/epoch after shift. - 68,799 notes, packed into 1,864 rows of 16,384 tokens with seed 731. - 59,996,020 scored supervised labels seen through this checkpoint. - No supervised trajectory examples. These are the notes-only controls, separate from notes + trajectory mixture models. - Eight GH200 GPUs, effective batch 8, learning rate… Open weights apache-2.0 9.7B parameters 262,144 tokens transformers [View model](https://savrn.com/models/qwen35-9b-harvey-v4-notes-ntp-30m) Model ### [qwen35-9b-multihop-attention-lora-ckpt536](https://savrn.com/models/qwen35-9b-multihop-attention-lora-ckpt536) [LeQ](https://savrn.com/model-publishers/ericq13) Open weights apache-2.0 [View model](https://savrn.com/models/qwen35-9b-multihop-attention-lora-ckpt536) Model ### [qwen35-9b-wmrl-v4-kl-mix30m](https://savrn.com/models/qwen35-9b-wmrl-v4-kl-mix30m) [Violet Xiang](https://savrn.com/model-publishers/violetxi) wm-internalization v4 checkpoint — condition kl-mix30m, save final. Full-finetune of Qwen/Qwen3.5-9B on the Calderwood & Harkness synthetic law-firm corpus (world-internalization study, v4 lineage: 9B student, ~50k think-on seed pool). Grafted back into the hub composite layout (Qwen35ForConditionalGeneration) — servable with vLLM out of the box. Open weights apache-2.0 9.7B parameters 262,144 tokens [View model](https://savrn.com/models/qwen35-9b-wmrl-v4-kl-mix30m) Model ### [qwen35_gui_gguf](https://savrn.com/models/qwen35-gui-gguf) [Jaden Dudik](https://savrn.com/model-publishers/itsnotacreativeuser) Open weights [View model](https://savrn.com/models/qwen35-gui-gguf) Model · Text generation ### [qwen35_gui_lora](https://savrn.com/models/qwen35-gui-lora) [Jaden Dudik](https://savrn.com/model-publishers/itsnotacreativeuser) Open weights peft [View model](https://savrn.com/models/qwen35-gui-lora) Model · Text generation ### [qwen35a3b-moe-expert-paging-q4](https://savrn.com/models/qwen35a3b-moe-expert-paging-q4) [Max Sun](https://savrn.com/model-publishers/hellosun) 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… Open weights apache-2.0 gguf [View model](https://savrn.com/models/qwen35a3b-moe-expert-paging-q4) Model ### [qwen35b-rho80-noshuffle-step_100-hf](https://savrn.com/models/qwen35b-rho80-noshuffle-step-100-hf) [Yu](https://savrn.com/model-publishers/tanyunvonnv) Open weights [View model](https://savrn.com/models/qwen35b-rho80-noshuffle-step-100-hf) Model ### [qwen35b-tailsft-sft-step100](https://savrn.com/models/qwen35b-tailsft-sft-step100) [Yu](https://savrn.com/model-publishers/tanyunvonnv) Open weights 36B parameters 262,144 tokens [View model](https://savrn.com/models/qwen35b-tailsft-sft-step100) Model ### [qwen35moe-mtp-head-GGUF](https://savrn.com/models/qwen35moe-mtp-head-gguf) [Yabau](https://savrn.com/model-publishers/yabau) Open weights llama.cpp [View model](https://savrn.com/models/qwen35moe-mtp-head-gguf) Model ### [qwen36-27b-sae2m-l42](https://savrn.com/models/qwen36-27b-sae2m-l42) [De Schamphelaere](https://savrn.com/model-publishers/ceselder) A dictionarylearning-format BatchTopK sparse autoencoder on the post-layer-42 residual stream of Qwen/Qwen3.6-27B (d = 5120), with F = 2,097,152 (2^21, 409.6x) features and k = 64. Companion to the 131,072-feature SAE at ceselder/qwen36-27b-sae-l42 (same layer, same recipe, 16x smaller dictionary). BatchTopK (Bussmann et al.), dictionarylearning-style trainer run feature-parallel over 8 GPUs (each rank owns 262,144 features; global top-k via all-gather, all-reduced reconstruction, aux-k loss alpha 1/32 with 40,960 aux features, learned activation threshold with EMA beta 0.999). lr 1.77e-5, warm-up 1,000 steps, decay from step 195,312, batch 4,096 tokens x 244,140 steps = 1.0e9 tokens… Open weights apache-2.0 [View model](https://savrn.com/models/qwen36-27b-sae2m-l42) Model ### [qwen38-27b-ap-d01-seed42](https://savrn.com/models/qwen38-27b-ap-d01-seed42) [David Chew](https://savrn.com/model-publishers/david-chew-hl) Uncensored build of Qwen/Qwen3.8-27B, produced with Apostate using the diode path. The diode repurposes one MLP neuron per layer into a gated refusal subtractor: it removes the residual refusal direction only when a benign-calibrated detector fires above threshold, so benign inputs keep the original weights. The result is a plain checkpoint: no runtime hook, adapter, finetune, or router. Delivery and KL are measured separately by apostate test, not during the bake; diodereport.json records the edit settings. This is a standard Transformers checkpoint. This model is uncensored and will answer harmful and dangerous requests. You are responsible for how you use it. Open weights 26.9B parameters 262,144 tokens [View model](https://savrn.com/models/qwen38-27b-ap-d01-seed42) Model · Image and text to text ### [qwen38-27b-swerebench-grpo-step8](https://savrn.com/models/qwen38-27b-swerebench-grpo-step8) [Haixin Nan](https://savrn.com/model-publishers/wanna0720) This is a Hugging Face safetensors export of the completed Step 8 full-parameter GRPO checkpoint. The language-model tensors are converted from the Miles/Megatron training checkpoint. The visual and MTP tensors, tokenizer, configuration, and processor assets come from Qwen/Qwen3.8-27B at revision The run trained on 1,586 SWE-rebench v2 Filtered tasks using independent mini-swe-agent rollouts. It used eight NVIDIA B300 GPUs (TP=2, CP=4), 32 prompt groups with eight rollouts each per optimizer step, global batch size 256, and learning rate 2e-6. This is the no-branch GRPO baseline checkpoint, not a branching-policy checkpoint. Source training-checkpoint completion marker SHA-256… Open weights apache-2.0 transformers [View model](https://savrn.com/models/qwen38-27b-swerebench-grpo-step8) Model ### [qwen38-27b-swerebench-grpo-step8-training-state](https://savrn.com/models/qwen38-27b-swerebench-grpo-step8-training-state) [Haixin Nan](https://savrn.com/model-publishers/wanna0720) This repository preserves the completed Step 8 checkpoint from a full-parameter GRPO run using a 1,586-task SWE-rebench v2 Filtered corpus. It is a Miles/Megatron distributed checkpoint, including model and optimizer state, metadata, and the dataset cursor. It is intended for research recovery and continued training. The model weights for direct inference are published separately as a Hugging Face safetensors export. The archived run used eight NVIDIA B300 GPUs with tensor parallelism 2 and context parallelism 4. Each step requested 32 prompt groups with eight rollouts per group (global batch size 256); the learning rate was 2e-6. Training used mini-swe-agent, independent sandboxes and… Open weights apache-2.0 [View model](https://savrn.com/models/qwen38-27b-swerebench-grpo-step8-training-state) Model · Text generation ### [qwen38-40b-prune](https://savrn.com/models/qwen38-40b-prune) [DavidB](https://savrn.com/model-publishers/davd-b01) One upstream MoE, six generations of masks, one bit-exact surgery, a rebuilt factual memory, and a three-stage post-training (SFT → SimPO → RLVR) — with every error I made and every idea I falsified along the way. This repository is the complete technical record of the project: the activation-instrumentation artifacts, the mask-generation methodology, the surgery code, the PLE (n-gram memory) rebuild, the trainers, and the full lab notebook. The champion weights live in sibling repos What I set out to do: take Qwen3.8-Flash-Next — an ~180B-class MoE (~130B of core network + 51B of n-gram memory table + 4B MTP) — and cut it by 75% of its experts (512 → 128) while conserving the maximum… Open weights apache-2.0 transformers [View model](https://savrn.com/models/qwen38-40b-prune) Model ### [qwen38-dspark-moe-256exp](https://savrn.com/models/qwen38-dspark-moe-256exp) [Yusheng Su](https://savrn.com/model-publishers/yushengsu) 5-layer DSpark MoE drafter for Qwen3.8-27B with 256 routed experts (width 512, top-10) and one shared expert (2048): Kan Wu's kan/moe-3-qwen38 SpecForge recipe with half the experts, trained from scratch for 3 epochs on RadixArk/Qwen3.8-27B-Regen-Mixture-v1 (global batch 256, lr 5e-4 constant) on one 4x GB300 node. Each finished epoch is a complete servable SGLang export in its own folder: Training curves: https://wandb.ai/milestraining/specforge-5layer-moe-0.5%2Aexperts/runs/qwen38-dspark-moe-256exp-3ep-gb300 Recipe and config: branch exp/qwen38-dspark-moe-256exp-3ep-gb300 of https://github.com/yushengsu-thu/SpecForge Open weights apache-2.0 [View model](https://savrn.com/models/qwen38-dspark-moe-256exp) Model ### [qwen38-dspark-moe-2layer](https://savrn.com/models/qwen38-dspark-moe-2layer) [Yusheng Su](https://savrn.com/model-publishers/yushengsu) 2-layer MoE DSpark draft models for Qwen3.8-27B: an ablation of Kan Wu's 5-layer Qwen3.8-27B DSpark MoE drafter (SpecForge branch kan/moe-3-qwen38) with the draft reduced from 5 to 2 decoder layers and nothing else changed (512 routed experts x 512 wide, top-10 softmax routing, sigmoid-gated shared expert of width 2048, router centering; 8.48B total / about 1.08B active parameters). One servable export per training epoch, each in its own folder (config.json, model.safetensors, tokenizer and a README with the training metrics of that checkpoint): Serving needs an SGLang build with the Qwen3MoeDSparkModel draft class (upstream SGLang 0.5.18 only ships the dense Qwen3DSparkModel). Download one… Open weights apache-2.0 [View model](https://savrn.com/models/qwen38-dspark-moe-2layer) Model ### [qwen38-dspark-moe-2layer-1ep](https://savrn.com/models/qwen38-dspark-moe-2layer-1ep) [Yusheng Su](https://savrn.com/model-publishers/yushengsu) Placeholder. The servable export of the 2-layer Qwen3.8-27B DSpark MoE drafter after epoch 1 will be uploaded here automatically when that checkpoint is written. Live training metrics: https://wandb.ai/milestraining/specforge-2%20layer%20moe Code and recipes: https://github.com/yushengsu-thu/SpecForge/tree/exp/qwen3.8-27b-dspark-moe-2layer Open weights apache-2.0 [View model](https://savrn.com/models/qwen38-dspark-moe-2layer-1ep) Model ### [qwen38-dspark-moe-2layer-2ep](https://savrn.com/models/qwen38-dspark-moe-2layer-2ep) [Yusheng Su](https://savrn.com/model-publishers/yushengsu) Placeholder. The servable export of the 2-layer Qwen3.8-27B DSpark MoE drafter after epoch 2 will be uploaded here automatically when that checkpoint is written. Live training metrics: https://wandb.ai/milestraining/specforge-2%20layer%20moe Code and recipes: https://github.com/yushengsu-thu/SpecForge/tree/exp/qwen3.8-27b-dspark-moe-2layer Open weights apache-2.0 [View model](https://savrn.com/models/qwen38-dspark-moe-2layer-2ep) ## Model Collections Hand-picked starting points, each with the reason it exists. Collection · 4 entries ### [Embedding models for retrieval](https://savrn.com/collections/embedding-models-for-retrieval) Sentence and document embedding models used to build retrieval systems. Dimension and sequence length matter more than size here, and both come from the publisher. [View collection](https://savrn.com/collections/embedding-models-for-retrieval) Collection · 4 entries ### [Models that fit on one accelerator](https://savrn.com/collections/models-that-run-on-one-accelerator) Models whose publisher-reported parameter count puts them within reach of a single accelerator at common precisions. Memory needed depends on precision and serving configuration, so treat the parameter count as the starting point, not the answer. [View collection](https://savrn.com/collections/models-that-run-on-one-accelerator) Collection · 6 entries ### [Open-weight text models worth knowing](https://savrn.com/collections/open-weight-text-models-worth-knowing) Widely used open-weight language models, chosen because each one is a distinct family rather than a variant of the one above it. Selection, not a ranking. [View collection](https://savrn.com/collections/open-weight-text-models-worth-knowing) Collection · 3 entries ### [Speech and audio models](https://savrn.com/collections/speech-and-audio) Recognition and synthesis models, grouped so the two directions are easy to compare. [View collection](https://savrn.com/collections/speech-and-audio) ## Open-Weight Models Explained ### What is an open-weight model? An AI model whose trained weights are published for anyone to download, so it can be run, tested and fine-tuned on hardware the user controls. ### Is an open-weight model the same as open source? Not always. Open weights means the trained model can be downloaded. Open source usually also means the training code and data are available and the license allows broad reuse. Many open-weight models release the weights only. ### Can I use an open-weight model commercially? It depends on the license. Apache 2.0 and MIT allow commercial use. Other licenses limit it, for example to non-commercial use or below a set number of users. Every model page here shows its license. ### How much memory does an open-weight model need? About two bytes per parameter at 16-bit precision, so a 7-billion-parameter model needs roughly 14 GB for its weights, plus memory for the context it processes. Each model page lists its parameter count and the size of its files. ## Related SAVRN Research The hub sits beside SAVRN's market data and infrastructure research: what models cost to run, and what it takes to run them. SAVRN Index ### [What open models cost to run](https://savrn.com/ai-index/pricing/open-models) The same open-weight model priced by every host that serves it, per million tokens. Research Hub ### [Data center trackers and maps](https://savrn.com/data-center-research) Moratoriums, permits, power, water and capital behind the facilities that run these models. Method ### [How the Model Hub is built](https://savrn.com/model-hub/methodology) Sources, evidence labels, refresh behaviour, and the limits of every comparison here.