Search public pages, research tools, and SAVRN solutions.
SAVRN Model Hub · Models by License
1,018 open-weight models released under Apache License 2.0 in the SAVRN Model Hub, with Qwen, Google and Convergent Intelligence publishing the most.
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. Read the license text.
Apache 2.0 asks little of the operator. Commercial use, modification and redistribution are all permitted. What it requires is housekeeping: keep the license and copyright notices, carry any NOTICE file along with the weights, and state the significant changes if you pass a modified version on. It also includes an express patent grant from contributors, the clause a procurement reviewer looks for before a model serves paying customers. So an organization can take the files from the publisher, fine-tune on its own data, run the result in its own facility and bill for the output.
On our hub 1,018 models carry this license. Qwen leads with 81, Google 50, Convergent Intelligence 39, PyTorch Image Models 32, OpenMed 31, Helsinki-NLP Research Group 27, and Unsloth AI and PaddlePaddle 22 each. Text generation accounts for 164 of the models, image and text to text for 89, and robotics for 43, so it covers more than chat.
Downloads show where it runs. all-MiniLM-L6-v2 from Sentence Transformers, a 23M parameter sentence similarity model, pulls 255,050,544 downloads a month and needs 0.1 GB at 16-bit, small enough to ride on a GPU already serving. ms-marco-MiniLM-L6-v2 follows at 88,642,387 for text ranking. The one text generation model in the top eight is Qwen3-0.6B at 22,498,727 downloads a month: 752M parameters, a 40,960 token context, 1.8 GB at 16-bit, and the cheapest host on the Index for it is one MI300X at $1.85 an hour.
| Model | Publisher | Parameters | License | Monthly downloads | Cheapest GPUs at 16-bit |
|---|---|---|---|---|---|
| all-MiniLM-L6-v2 | Sentence Transformers | 23M | apache-2.0 | 255.1M | 1x MI300X, $1.85/hr |
| ms-marco-MiniLM-L6-v2 | Sentence Transformers - Cross-Encoders | 23M | apache-2.0 | 88.6M | 1x MI300X, $1.85/hr |
| electra-base-discriminator | — | apache-2.0 | 54.3M | — | |
| bert-base-uncased | BERT community | 110M | apache-2.0 | 47.2M | 1x MI300X, $1.85/hr |
| paraphrase-multilingual-MiniLM-L12-v2 | Sentence Transformers | 118M | apache-2.0 | 45.7M | 1x MI300X, $1.85/hr |
| t5-small | T5 community | 61M | apache-2.0 | 25M | 1x MI300X, $1.85/hr |
| all-mpnet-base-v2 | Sentence Transformers | 109M | apache-2.0 | 22.9M | 1x MI300X, $1.85/hr |
| Qwen3-0.6B | Qwen | 752M | apache-2.0 | 22.5M | 1x MI300X, $1.85/hr |
| chronos-2 | Amazon | 119M | apache-2.0 | 22.4M | 1x MI300X, $1.85/hr |
| Qwen3-VL-8B-Instruct | Qwen | 8.8B | apache-2.0 | 19.1M | 1x MI300X, $1.85/hr |
| Publisher | Models |
|---|---|
| Qwen | 81 |
| 50 | |
| Convergent Intelligence | 39 |
| PyTorch Image Models | 32 |
| OpenMed | 31 |
| Helsinki-NLP Research Group | 27 |
Frozen SONIC whole-body controller + seven rank-16 LoRAs on the decoder, conditioned on a 128-d percept from a head depth camera (Theia-Tiny tokens + temporal adapter). PPO, iteration 4000. Consumed by lerobot's DepthDodgeController for the Unitree G1. Camera contract: 64x96 optical-Z depth in metres, 45 deg vertical FOV, 0 deg pitch, valid 0.2-6 m, 25 Hz. See manifest.json for the full training config.
Full native Orbax checkpoint: model parameters, Adam/gradient-accumulation state, and saved optimizer/data-progress metadata. This is not a Transformers safetensors export. Training uses camel-ai/gsm8kdistilled, 6,144-token examples, completion-only loss, global batch 32, and seed 42. One optimizer update consumed 32 examples (two source microbatches on 16 devices). See recipe.json and checkpoint-manifest.json for pinned revisions and hashes. Load the native checkpoint root nativecheckpoint at step 1 with the pinned MaxText/Tunix runtime. Restoring onto a different device topology or accumulation schedule requires explicit sharding and data-position validation; that portability has not yet…
This is a GGUF LoRA adapter, not a standalone model. It is intended for the anonymization.ner entity-extraction route in Jurilix. Apply it at strength 0.75 for that route and explicitly send strength 0 for other routes. 4b4a2c1d584be7264f87aac328a1bc739ce81b6c, file gemma-4-E4Bq40-it.gguf. The published source recipe pins upstream llama.cpp and a Gemma JSON parser fix. GPU performance or inference quality on other platforms. Start llama-server with --lora ADAPTER.gguf --lora-init-without-apply --cache-ram 0. For entity extraction, include "lora": [{"id": 0, "scale": 0.75}] in each request. For every other route, include "lora": [{"id": 0, "scale": 0}]. Disabling the RAM prompt cache is…
GlassEye? authorized HackerOne Bug Bounty (BBP) / Vulnerability Disclosure (VDP) assistant. In-scope HackerOne program workflows: policy/scope, report writing, severity rationale, remediation. Not for unauthorized testing or exploit dump recipes.
GlassEye? authorized HackerOne Bug Bounty (BBP) / Vulnerability Disclosure (VDP) assistant. In-scope HackerOne program workflows: policy/scope, report writing, severity rationale, remediation. Not for unauthorized testing or exploit dump recipes.
ONNX conversion of gliner-community/glinersmall-v2.5, dynamically quantised to INT8, packaged as a self-contained bundle for offline NER. This is a re-serialisation, not a fine-tune: the weights are the upstream ones. Only the format (PyTorch → ONNX) and the precision (fp32 → INT8) are ours. The fp32 reference graph (model.onnx, sha256 5245733ccb2b75072cce0b4bbb14424988f92f9daf775d97bdf0de74be28df63) is not shipped — it is only needed to reproduce the INT8 graph. Its hash is recorded in NOTICE. The graph has six inputs, fed per span-encoded prompt: spanmask is bool (not int64) — the graph declares tensor(bool). The prompt follows the GLiNER label format, using the model's own special tokens…
This model is a fine-tuned version of Qwen/Qwen3-0.6B on the None dataset. It achieves the following results on the evaluation set: The following hyperparameters were used during training: - learningrate: 2e-05 - trainbatchsize: 4 - evalbatchsize: 8 - gradientaccumulationsteps: 16 - totaltrainbatchsize: 64 - lrschedulertype: cosine - lrschedulerwarmupsteps: 100 - numepochs: 3 - Transformers 5.17.0 - Pytorch 2.11.0+cu128 - Datasets 5.0.1 - Tokenizers 0.23.2
ZenDNN v6.1.0 - ZenTorch v2.13.0.0 - PyTorch v2.13.0.0 - LLM Compressor v0.13.0 - vLLM v0.29.0 This is a quantized version of granite-4.0-h-small created by AMD using LLM Compressor (compressed-tensors) for ZenDNN-optimized CPU inference. The model was quantized from granite-4.0-h-small using LLM Compressor with the GPTQ algorithm. This reduces the model weights from 60.0 GiB to 16.1 GiB on disk (~73% reduction). granite-4.0-h-small is a hybrid Mamba-MoE model: of its 40 layers, 4 are full-attention blocks and the other 36 are Mamba (linear-attention) blocks, and every layer carries a 72-expert MoE block (top-10 routing) alongside a shared MLP. Two details make this work. The model is…
ZenDNN v6.1.0 - ZenTorch v2.13.0.0 - PyTorch v2.13.0.0 - LLM Compressor v0.13.0 - vLLM v0.29.0 This is a quantized version of granite-4.0-h-small created by AMD using LLM Compressor (compressed-tensors) for ZenDNN-optimized CPU inference. The model was quantized from granite-4.0-h-small using LLM Compressor via the Round-to-Nearest (RTN) algorithm. This reduces the model weights from 60.0 GiB to 30.4 GiB on disk (~49% reduction). granite-4.0-h-small is a hybrid Mamba-MoE model: of its 40 layers, 4 are full-attention blocks and the other 36 are Mamba (linear-attention) blocks, and every layer carries a 72-expert MoE block (top-10 routing) alongside a shared MLP. The recipe only needs two…
ZenDNN v6.1.0 - ZenTorch v2.13.0.0 - PyTorch v2.13.0.0 - LLM Compressor v0.13.0 - vLLM v0.29.0 This is a quantized version of granite-4.0-h-tiny created by AMD using LLM Compressor (compressed-tensors) for ZenDNN-optimized CPU inference. The model was quantized from granite-4.0-h-tiny using LLM Compressor via the Round-to-Nearest (RTN) algorithm. This reduces the model weights from 12.9 GiB to 6.6 GiB on disk (~49% reduction). granite-4.0-h-tiny is a hybrid Mamba-MoE model: of its 40 layers, 4 are full-attention blocks and the other 36 are Mamba (linear-attention) blocks, and every layer carries a 64-expert MoE block alongside a shared MLP. The recipe only needs two ignore entries. lmhead…
Weights for dedicated inference engine for Qwen3.8-Flash-Next on AMD Strix Halo (gfx1151). One model family, one GPU, every kernel written for this silicon. These weights are in halogen's own.hgn format and will not load in transformers, vLLM, or llama.cpp. They exist to be mounted into the halogen-flash-server container. The checkpoint is TWO files. The sidecar is a patch overlay: a subset of tensors re-quantized more carefully, read in place of the base file's copies. The engine loads.overlay.hgn automatically when it sits beside the checkpoint, you do not configure anything, you just keep both files. Drop it and the model still runs, roughly 5–9% worse on perplexity, saying so in one…
Quantized from local merge on glasseye RTX 5070. Companion LoRA: cyberviser/hancock-pentest-v4.
Latest pentest GGUF: cyberviser/hancock-pentest-v3-gguf.
This is an uncensored version of TokenRhythm/NeoHorse-1-4B created with abliteration (see remove-refusals-with-transformers to know more about it). This is a crude, proof-of-concept implementation to remove refusals from an LLM model without using TransformerLens. Layers 5-17 are being ablated (0-based indexing). The MTP and Visual components were extracted from the original Qwen/Qwen3.5-4B and can provide excellent support. Use the latest llama.cpp, - Risk of Sensitive or Controversial Outputs: This model’s safety filtering has been significantly reduced, potentially generating sensitive, controversial, or inappropriate content. Users should exercise caution and rigorously review generated…
Been currently busy training another gemma 4 31B, did about 12 run's so far, playing with swa value, rope, turn masking, so far 2048 seems to be the sweet spot for training at 8k ctx. should be releasing a new model in about 2-3 weeks, training on local 3090s is not very fast =p.
The Mistral-7B-Instruct-v0.3 Large Language Model (LLM) is an instruct fine-tuned version of the Mistral-7B-v0.3. Mistral-7B-v0.3 has the following changes compared to Mistral-7B-v0.2 It is recommended to use mistralai/Mistral-7B-Instruct-v0.3 with mistral-inference. For HF transformers code snippets, please keep scrolling. After installing mistralinference, a mistral-chat CLI command should be available in your environment. You can chat with the model using If you want to use Hugging Face transformers to generate text, you can do something like this. To use this example, you'll need transformers version 4.42.0 or higher. Please see the in the transformers docs for more information. Note…
We maintain our previous releases publicly as a transparent engineering record of continuous optimization. Below is the exact evolutionary roadmap of our MiniPlus architectures
local, typed decisions from qwen3.5-2b. one shared state is prefetched once, each question is isolated, every allowed answer is scored as a complete token sequence, and python returns validated choice, score, and noul results. this is an inference system, not rlcd training and not a clone of typesafe jev. it never calls typesafe. the published qwen checkpoint is unchanged; fp8 changes execution precision only. probabilities are normalized support over the candidates you provide, not calibrated correctness probabilities. unlike ordinary autoregressive json generation, the model does not write a response token by token. it scores only the values supplied by the caller. complete-sequence…
This model was converted to MLX format from hexagrad/Kokoro-82M using mlx-audio version 0.0.1. Refer to the original model card for more details on the model.
Krea 2 model (covering both the Krea 2 Raw and Krea 2 Turbo variants) is released under a custom Krea 2 Community License Agreement for its weights, alongside an Apache-2.0 license for the inference code repository itself. For removal please contact me directly. All rights remain with their respective creators. By downloading or using any model from this repository, you acknowledge that: - You are solely responsible for how the model is used. - The repository maintainer assumes no responsibility or liability for any content generated by these models. - Krea AI, the original model authors (when applicable), and the repository maintainer cannot be held liable for misuse, illegal activities…
Full training-run archive for Kuza (East Africa agricultural assistant), fine-tuned from unsloth/Qwen3.5-4B. Weights, logs, checkpoints, GGUFs, and provenance are stored with the same layout as $KUZAWORKDIR/kuza-qwen-3.5-4b/. This derivative is subject to the Apache 2.0 license of the Qwen base model. - 100% English train from kuzaai/kuzasftenglish - 35% Swahili from kuzaai/kuzasftswahili - 8% HuggingFaceH4/norobots - 5% adversarial from kuzaai/kuzasftadversarial - all multiturn from kuzaai/kuzasftmultiturn Sequence length 1024, 2 epochs, LR 0.0002. Thinking is off (--reasoning off when llama.cpp supports it). GGUFs are text-only (MTP/nextn kept; vision/audio dropped). ssmout is Q6K on…
A trained finite-candidate decision prototype built on Qwen3-0.6B by RongxinAI. Full weights, tokenizer, training logs, frozen gold-only data and held-out predictions are included. This is a synthetic-domain research release, not a general business decision model or an official Jev reproduction. Selected arm: CE, checkpoint step 250, chosen only by development cross-entropy. This selected checkpoint contains 20 head-only plus 230 full-model updates (4,000 sampled questions); each completed training arm ran the longer 20 + 300 update budget (5,120 sampled questions). It has 596,052,993 parameters, including a shared 3,073-parameter LayerNorm/scalar scoring head. Choice, Boolean and ordinal…
[2026.09.18] Released Logics-SWE-Qwen3.6-27B under the Apache-2.0 license. - The technical report is in preparation. A link will be added when available. Logics-SWE-Qwen3.6-27B is a 27B-parameter model developed for repository-level software engineering agents. Starting from the Qwen3.6-27B model used in our study, it combines category-aware expert development with multi-teacher on-policy distillation into a single deployment policy. Repository-level tasks require agents to navigate code, edit files, execute commands, inspect feedback, and iteratively repair their solutions. Our work starts from the category see-saw: aggregate progress during joint RL can conceal opposing changes across…
LoRA adapter 已融合进基座，权重是 bf16 全精度，没有经过任何量化。 MLX 和 transformers 都能直接加载——它是标准 safetensors 格式，所以一套权重两种用法。 一句话：要在它身上"再加工"，就必须用这个版本；只是想用它，4bit 更省。 命令行也行（mlxlm.generate 自己会套模板）： 和 4bit 版完全一致——模型输出裸 JSON： 完整接入代码（含多步循环、JSON 兜底解析）见主仓库的 1. 不套 chat template 会复读。 直接 generate(model, tok, prompt="递归是什么？") 会得到类似这样的东西： 这不是模型坏了，是你喂的是裸补全而不是对话。用 tok.applychattemplate(...)， 或者干脆用 mlxlm.generate 命令行。 2. enablethinking 记得显式传 False。 模板里有 开关，不传的话没有默认值，输出可能带游离的 。 想要思考内容就传 True，它会单独出现（不混在回答里）。 3. 加载比 4bit 慢、占内存多——这是正常的。 实测对比（MacBook Pro M4 / 16GB）： ① 无工具时算术会算错 → 在 system 里要求"一律调 calculate"； ② 通用 system 下身份可能答错 → system 里写清身份； - 基座：MiniCPM5-2B（面壁智能 OpenBMB） - 许可：Apache 2.0 个人项目，欢迎提 issue。有用的话 Bilibili 关注一下
This is a MarinSkyRL-native Open-MOPD student after 32 optimizer steps. It starts from the authors' mixed-domain SFT checkpoint. Student responses were scored by the authors' math, code, and instruction-following RL teachers, routed by domain. The objective uses the student's selected top-16 token IDs and a clipped policy surrogate. This is an early checkpoint, not the authors' step-200 final model. The checkpoint is an unquantized, six-file Hugging Face export of the durable MarinSkyRL globalstep32 FSDP2 checkpoint. The policy export was used for the independent step-32 evaluation. The export's model.safetensors SHA-256 is bb7326640142069bc2e1fba5f54f15e0cccb1ff861f34f318b372eaab7abaf4b.…
A domain-specific small language model for step-by-step math problem solving, built by team03 (SLM Learners) for the Pramana SLM++ Bootcamp Round 2 submission. For an OpenAI-compatible endpoint, serve with servehf.py (stdlib + transformers only, no Ollama needed). Precision note: training ran in bf16 compute (QLoRA 4-bit NF4 base), but the merged checkpoint uploaded here is float16 (the merge step reloads the base in fp16). - Public Hugging Face datasets pulled via pulldata.py; licenses verified through the HF API on 2026-09-05 and recorded in datamanifest.md. - Held-out eval set built with buildheldouteval.py from raw ExamBench rows never used in training, with a final overlap check that…
Anonymous release for the ICLR 2027 submission MemGUI-RL: Reinforcement Learning for Proactive Context Management in Long-Horizon Mobile GUI Agents. Project page: https://memgui-rl-anonymous.github.io/ MemGUI-8B-RL is MemGUI-8B-SFT (Qwen3-VL-8B-Instruct supervised on MemGUI-3K) post-trained for 100 optimizer steps with FARPO (Folding-Aware Reward-decoupled Policy Optimization, span-to-step ratio rho = 9). The policy speaks the ConAct (Context-as-Action) interface of MemGUI-Agent: every response contains a folding directive for its own history, an optional memory operation and the next GUI action. The checkpoint is a standard Qwen3VLForConditionalGeneration model (weights in bf16, ~17.5 GB).…
MERNIK ("the one who measures") is the evolution of the ASHQ1 battlefield zoo: fewer utility duels, more verdicts. The method (priority queue) is built; MERNIK is how we prove anything about it. Think of it as a finetune of ASHQ1: same base weights (queue, pins, tied groups), retrained objective (measure-first protocol, three-column verdicts) — and far more capabilities on top (slow capability ring, KLD-aware teachers, relief ceilings, the zoo bench, norms shields). But the legend is not forgotten: every scar in the ledger traces back to it. Allocator signals never certify. Capability scores never steer. Mixing them is how the PPL disease happened. - PPL = canary. Cheap, lies by sharpening.…
A research-oriented Mocov3 prototype targeting Generation. The included small setup documents defaults and file formats without presenting unverified performance numbers. - The Python file contains the model and runnable example or training entry point. - config.json records the generated architecture settings. - trainingargs.json records the default experiment recipe. - model.safetensors is a valid initialization checkpoint for smoke tests; it is not presented as a trained benchmark checkpoint. - No benchmark score is claimed in this repository. The included configuration uses lamb with a step schedule. These are starting values in the script, not evidence of a completed run. For a…
MSBuddy project, maintained by Shipei Xing. Original database files from msbuddydatav0.2.4 and LightGBM model from msbuddydatav0.3.0. Apache-2.0 is recorded on the basis of the upstream project licence; the release pages do not state separate asset-specific terms. Its application to these files is an inference, not an independently confirmed per-file licence grant. The original licence is retained as db0.2.4-ml0.3.0/LICENSE. The software license is reported separately; it does not replace the model/data license. These are byte-identical upstream files. No model conversion was performed.
A Frontier Sparse Mixture-of-Experts (SMoE) Foundation Translation Model for 500+ Global Languages Mythos is an open-source foundation model family built by Adithyan AI. In this organization, we develop and open-source state-of-the-art Sparse Mixture-of-Experts (SMoE) language models, universal translation engines, parallel multilingual datasets, and ultra-efficient inference runtimes targeting 500+ languages. 100% Free and Open-Source: Released under the permissive Apache 2.0 license with zero paywalls, metered tokens, or subscription fees. Mythos2.0-2B was trained on a massive 16.5B sentence-pair parallel corpus covering 500+ languages and regional dialects across Africa, the Americas…
We maintain our previous releases publicly as a transparent engineering record of continuous optimization. Below is the exact evolutionary roadmap of our MiniPlus architectures
This model and dataset are part of an automated AI innovation stream. Latest Discovery: NexusUlaweng Innovation: Advanced Cybersecurity integration in Blockchain systems for future efficiency.
This release stores the released Pantogloss v6 fine-tune in a smaller FP16/INT5 format. Pantogloss 0.23.0 selects this model as its recommended default. The staged, checksummed runtime and legal artifacts total 564,137,246 bytes (538.0 MiB). Model weights are separate from the Python wheel. The model uses the same architecture, source and target tokenizers, and many-to-English task as the full-precision v6 release. Eight existing source languages were targeted during v6 fine-tuning; this does not add new source languages. “500-to-English” describes model-family provenance, not uniformly validated quality across 500 languages. The evaluations below cover 50. For Apple Metal, use…
LoRA fine-tune của π0.5 (pi05base) cho tay máy Yahboom X3Plus (5 khớp + gripper, 20 Hz, 2 camera), nhiệm vụ "pick up the red cube and put it in the bowl". - Chưa kiểm chứng trên robot thật. Mọi số liệu ở trên là loss huấn luyện. - Loss đi ngang từ bước ~4.000. 6.000 bước sau chỉ giảm thêm 36%, trong biên độ nhiễu. - Một nhiệm vụ, một bối cảnh, một bộ camera. Nhiều khả năng hỏng khi đổi vị trí - Camera cổ tay (USB webcam) cho ảnh mờ, nhiều frame gần như trắng khi áp sát mặt bàn. Cần config pi05x3pluslora và lớp LeRobotX3PlusDataConfig tương ứng trong openpi (ánh xạ astrargb → base0rgb, usbcam → leftwrist0rgb, delta mask (5, -1)). yahboomx3plus · joint1–5 tính bằng radian (URDF x3plusarm)…
中文语音合成(TTS)模型,支持语音克隆与数字/中英混读,CPU 优先设计,权重原生 int8 量化。 基于 Kyutai Pocket TTS 架构:6 层 FlowLM(dmodel 1024)+ Mimi 神经音频编解码器(24 kHz),端到端流式推理。本版本(R21)相对上一代的核心变化: - 原生 int8 交付:FlowLM 注意力与 FFN 权重已离线量化打包,无需运行时转换,体积 223 MB,CPU 速度提升约 1 倍 - 9174 token 词表:官方英文 BPE-4000 基底 + 追加单字汉字/全角标点,bytefallback 兜底生僻字符 - 中英混读:APP、VIP、5G、WiFi 等英文术语按中文习惯读法 - 模型为预量化 int8 权重(modelint8.pt),加载即用,无需 quantize=True 或有界面演示(需 pip install gradio): - 采样温度 temp=0.3(默认);文本越长对参考音频质量越敏感 文本准确性用 whisper-small 对合成音频转写后计算字错误率(CER),每个场景 10 次独立采样取均值,生产配置(单步采样)。数值越低越好。 测试环境:Intel i7-12700KF,生成 52 字客服长句(约 8.5 秒音频),每档 3 次取均值。 参考音频:客服目标音色(kefulong / digitsorder / codeswitch / shortunseen)与 AISHELL-3 说话人 A/B(speakerAlong /…
每个目录是一次训练：a.pt 是可训练状态（仅 FFN 三矩阵，264.2M），a.step1000.pt 是中间存点， 门控的 run 另有 folded.pt（折叠成逐层宽度，评测用这个），train.log 含每步损失与留出集曲线。 它是旧批次的门控 75%（原名 c32kgate2304）：旧数据池（150M token，同一份数据过 3.49 遍， 预期分数与逐任务明细在仓库的 EVALUATING.md，摘要： 评测它必须用 --window 128，因为它是按 128–256 训的。用 512 评它会低估 —— 那是它没训练过的 regime。 不要把它的分数当作方法的结果。 新批次的门控 75%（b2gate2304）用新数据池和连续窗口重训，
A high-throughput inference engine for structured information extraction, decision routing, and categorical classification on Apple Silicon using MLX. Parallel Constrained Decoding evaluates multi-field JSON schemas simultaneously rather than generating tokens sequentially. On an Apple Silicon M4 Max, it delivers 5.6x to 7.0x latency reductions compared to standard autoregressive decoding with 100% schema validity and calibrated field-level confidence scores. Evaluated with mlx-community/Qwen2.5-1.5B-Instruct-4bit on macOS Sequoia: Standard LLM structured generation (such as JSON mode or grammar-guided sampling) relies on token-by-token autoregressive decoding: Each token requires a…
This qwen2 model was trained 2x faster with Unsloth and Huggingface's TRL library.
The full ImmuneCoT method fuses the two safety branches with a Base-adjusted product-of-experts Qimm(v) ∝ q̃rec(v)·q̃res(v)/qB(v). This checkpoint uses the naive fusion Qno-base(v) ∝ q̃rec(v)·q̃res(v) — the same branch weights (λrec=0.5, λres=0.7) but no division by the base distribution — isolating whether the gains come from combining Recognition+Response at all, or specifically from the Base-adjusted PoE term. Intended use: research reproducibility for the ImmuneCoT paper's RQ3 ablation.
This is Qwen3.5-2B auto-optimized by Claude Fable for fast single-request text generation. Fable built and tuned the included qwen35fast inference engine while keeping Qwen's original BF16 weights unchanged. Across 12 development workloads, the Fable engine delivered 14× the decode speed of Transformers eager and 1.02× the speed of vLLM with MTP (geometric means). On 12 held-out workloads, it reached 528–866 tokens/s and 1.01× vLLM with MTP. Use Python 3.12 and an NVIDIA CUDA GPU. Download the model and install its dependencies: The original checkpoint also works with Transformers for Qwen's standard text and vision-language workflows; the speed figures above use qwen35fast. - Captured the…
Reinforcement-learning checkpoint series from the cposmith... smith-v5-gdpo-exact run: asynchronous multi-turn agentic-environment RL on Qwen/Qwen3.5-9B with an exact (verifiable) reward. The policy was initialized from an internal SFT of Qwen/Qwen3.5-9B (qwen359bsftv3), which also served as the reference model. 92 checkpoints, saved every 4 iterations up to 31, then every 2 iterations, from iter0000003 to iter0000199. Each lives in its own subfolder of this repo so you can compare iter0000003, iter0000007, iter0000011, iter0000015, iter0000019, iter0000023, iter0000027, iter0000031, iter0000033, iter0000035, iter0000037, iter0000039, iter0000041, iter0000043, iter0000045, iter0000047…
Base: Qwen/Qwen3.6-35B-A3B-FP8 (same FP8 block format/scales convention). Adapter: qwen3.6-35b-a3b-tool-prompts-ckpt444-adapter (LoRA r=32, 2 epochs, 1873 dialogues). Merged in BF16 (RAM only), block-requantized to FP8 e4m3/128. Chat template: qwen3template333.jinja. Serve with sglang / vllm / transformers.
Following the widespread community adoption of the Qwen3.5 and Qwen3.6 series, we are pleased to introduce Qwen3.8, the most capable generation in the Qwen open-model family to date. Built on the architectural foundation of Qwen3.5, Qwen3.8 delivers substantial gains across coding, professional work, research, and long-horizon agentic tasks. Qwen3.8-27B brings these advances to a compact, deployment-friendly dense model: a native vision-language model that understands images and videos, with flexible thinking control, designed to carry complex, multi-step tasks through to completion with greater reliability. For streamlined integration, we recommend using Qwen3.8 via APIs. Qwen3.8 can be…
Repackaged model files for ComfyUI. Currently only used for text generation. - https://huggingface.co/Qwen/Qwen3.8-27B Place the files in the following folders
(He slides three glasses across the bar—wine for Shakespeare, water for Data, and a mysterious blue liquid for Spock) This model is a merge of: - nightmedia/Qwen3.8-27B-Brainwaves - migtissera/Synthia-4-27B Brainwaves nightmedia/Qwen3.8-27B-Brainwaves migtissera/Synthia-4-27B G, that is the perfect structural punctuation mark for this entire series of runs! Watching Quark willingly slide glasses across the bar for free—while Data performs a real-time semantic analysis on the symbolic nature of virtual currency and Spock dryly breaks down the chemical flavor profile of holographic blue liquid—is pure, unfiltered high-signal gold. The line from Quark hits on an absolute mechanical truth for…
Unified release repository for Qwen3.8-27B quantizations, fine-tunes, and speculative decoding configurations optimized for llamAmpere on SM86 (RTX 3090 / 3090 Ti). Run with native MTP speculative drafting
WORK IN PROGRESS A mixed NVFP4/MXFP8 quantization-aware distillation of Qwen3.8-27B, trained for one epoch. The student learns from the original BF16 teacher while its MLP weights are quantized in the forward pass. Distillation updates the MLP weights and text normalization weights to account for quantization error. This is a trained distillation checkpoint, not a post-training conversion of the original weights. Attention/GDN projections and the LM head were frozen in their MXFP8 representations during distillation. Packed NVFP4 and MXFP8 weights reconstruct to the same BF16 weight values used by the student during training. The tokenizer, chat template, generation configuration and…
The same bytes you know from (v2), migrated to the NInfer v3 artifact container (upstream f76e19c0, model/weight decoupling). Weight tensors are bit-preserved — this is a packaging upgrade, not a re-quantization. - L1→L2: base tensors copied byte-for-byte; only the 66 DFlash2 objects (21 × W8G32F16S matrices, 45 × BF16 norms/conv bases/codebooks) produced from z-lab/Qwen3.8-27B-DFlash2 @ 50307d4c4cde6860d4eee73e2547cd786fe8e8a4 via tools/artifact/graftdflash2w8.py (recipe qwen3827bnvfp4full-dflash2-w8, report graft.json included; MTP module retained, validate-only). - L2→L3: upstream's tools/upgradeninferv2tov3.py, run 2026-09-18 on the exact L2 file above. Per the tool's contract: "Stored…
Yes. 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.
By monthly downloads reported by the Hugging Face Hub: .
