Users (both direct and downstream) should be made aware of the risks, biases and limitations of the model. Use the code below to get started with the model. Stage 1 Stage 2 Run04 used the original text-based SFT training configuration, not completion-only SFT. Carbon emissions can be estimated using the Machine Learning Impact calculator presented in Lacoste et al. (2019). This model was fine-tuned using Qwen3-4B-Base and the Tulu 3 and GSM8K datasets. Please cite the original works when using or referencing this model. Qwen3 title={Qwen3 Technical Report}, author={Yang, An and others}, year={2025} Tulu 3 title={Tulu 3: Pushing Frontiers in Open Language Model Post-Training}…
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.
11,734Models
3,323Datasets
288Papers
7,150Publishers
12,282Sourced relationships
Updated 2026-10-20 · How the library is built
11,734 models, sorted by most downloaded.
Qwen/Qwen3-4B-Base Users (both direct and downstream) should be made aware of the risks, biases and limitations of the model. Use the code below to get started with the model. This adapter was originally trained in an earlier run and reused unchanged for Run04. Carbon emissions can be estimated using the Machine Learning Impact calculator presented in Lacoste et al. (2019). This model was fine-tuned using Qwen3-4B-Base and the Tulu 3 and GSM8K datasets. Please cite the original works when using or referencing this model. Qwen3 title={Qwen3 Technical Report}, author={Yang, An and others}, year={2025} Tulu 3 title={Tulu 3: Pushing Frontiers in Open Language Model Post-Training}…
GENIAC-PRIZE 2026 テーマ2 コンペスクリーニング（学生・人材育成枠）の参加者が作成した開発モデルです。 ベースモデル Qwen/Qwen3-4B-Base に 2 段階の QLoRA SFT（事後学習）を施し、LoRA をマージしたものです。 - ライセンス: apache-2.0（ベースモデルの条件に従います） - Qwen/Qwen3-4B-Base (https://huggingface.co/Qwen/Qwen3-4B-Base) - 生成元の問題はすべて上記 PRM800K の train split（12,000 行）から機械的に層化抽出したものです（評価用データは一切使用していません）。 - Qwen/Qwen3-4B-Base 自身にその問題を解かせ、出力中の \boxed{...} の内容が参照解と正規化一致したものだけを機械的なルールのみで採用しました（1,698 件）。 - 人手による作成・加工・選別、および他のモデル・API の利用は行っていません。 - GSM8K / LIMO / Aya / OpenAssistant / Tulu 3 Persona IF は使用していません。 - したがって Tulu 3 Persona IF (ODC-BY-1.0) の帰属表示は不要です。 2 段階の QLoRA SFT。データの取得・整形・学習はすべて Google Colab 上で実行しました。 - maxseqlength 3072、bf16、targetmodules = q, k, v, o, gate, up, down…
集群下线前的离线备份，2026-09-14。共 73 个训练目录、23.9 GB，全部是 PEFT LoRA adapter（没有全参权重）。2026-09-16 补了 geoworldloopfull256k 与 geoworldloopm17full256k，09-17 补了 geoworldloopbd8full256k、geoworldloopr17full256k 与 geoworldloopr17lfull256k，09-18 补了 headtrain 三臂 geoworldloopr17l{1,2,3}hfull256k，09-19 补了 geoworldloopr17lsfull256k，09-21 补了 geoworldloopr17l{1,2}hsfull256k。目录结构与原仓库的 checkpoints/ 一比一相同，所以 src/interp/common.py 里 CARRIERS 的 adapter=ROOT / "checkpoints/ /final" 可以直接指过来。 代码与报告在 GitHub 仓库 geography（分支 cleantrn）；本仓库只补上被.gitignore 挡掉的权重。 单个载体（例如 TRN-geodesic 的 worldfar）： 这个 adapter 是带着注意力硬约束训练的：题面按 token 分成前缀、位移（距离与方位角）、城名、题尾四段， 第 0 到 16 层里答案侧（:）看不到城名、城名侧全程看不到距离与方位角。用它做推理必须挂同一套 mask （GitHub 仓库 geography 的…
Experiments on steering Anima, a 2B illustration model built on NVIDIA Cosmos-Predict2, with a second conditioning source: Beatrix, a byte-level language model from the geolip line. Anima reads its prompt through a small language model (Qwen3 0.6B) and a light adapter, so an added signal is not drowned by a very large text encoder; that makes it a bed for testing how far a second source can steer the image without a full diffusion training run. The first experiments measure the bed itself: how its conditioning responds to mood words and to a mood direction added to it, and what LoRAs trained on the model's own mood images do. Each experiment has its own folder under experiments/ with a…
100x faster than generative LLMs • Runs on laptops & cloud CPUs • Global #1 on JevBench When you ask ChatGPT or Claude a question, it generates words one token at a time, like a person typing out an essay. That takes 2 to 5 seconds and burns expensive GPU compute. That is great for writing a story, but it is painfully slow and expensive for simple decisions: - "Did the AI make up this answer, or is it actually in the PDF?" - "Should this customer's message go to billing, shipping, or technical support?" - "Does the revenue bar chart support this financial claim?" - "Did the student get the math problem right according to the answer key?" Psychologist Daniel Kahneman described human thinking…
100x faster than generative LLMs • Runs on laptops & cloud CPUs • Global #1 on JevBench When you ask ChatGPT or Claude a question, it generates words one token at a time, like a person typing out an essay. That takes 2 to 5 seconds and burns expensive GPU compute. That is great for writing a story, but it is painfully slow and expensive for simple decisions: - "Did the AI make up this answer, or is it actually in the PDF?" - "Should this customer's message go to billing, shipping, or technical support?" - "Does the revenue bar chart support this financial claim?" - "Did the student get the math problem right according to the answer key?" Psychologist Daniel Kahneman described human thinking…
End-to-end multi-lingual Variational Inference Text-to-Speech (VITS2) model covering 41 indigenous Ghanaian languages plus Ghanaian English. Checkpoints are automatically synced live from the NVIDIA H200 VM: - bertvits2checkpoints/: Current production Bert-VITS2 checkpoints (G.pth, D.pth, config.json). - v1autoregressivecheckpoints/: Previous discrete autoregressive experiment checkpoints.
Voice conversion for Asante Twi (twi). An audio-to-audio model that converts speech into the Zephyr voice while preserving the language. This checkpoint is dedicated to Asante Twi: it was fine-tuned only on Asante Twi (twi) clips from starting from the zero-shot seed-uvit-whisper-small-wavenet base. Use it to convert audio in Asante Twi. For other languages — or as a cross-lingual fallback — use ghanaopenai/ghana-vc. The library auto-selects this model when the dataset/config name matches.
Voice conversion for Avatime (avn). An audio-to-audio model that converts speech into the Zephyr voice while preserving the language. This checkpoint is dedicated to Avatime: it was fine-tuned only on Avatime (avn) clips from starting from the zero-shot seed-uvit-whisper-small-wavenet base. Use it to convert audio in Avatime. For other languages — or as a cross-lingual fallback — use ghanaopenai/ghana-vc. The library auto-selects this model when the dataset/config name matches.
Predicts where a league-average goalkeeper would position themselves, including the in-possession high-sweeper regime that the standard Ghost-GK v1 model cannot represent. Point-estimate (ghostgkx/y) served as the exact pickle-free boosted HistGradientBoostingRegressor mean. Part of the silly-kicks soccer analytics library (TF-60 rest-defense; ADR-083). The bundled Ghost-GK default/full variants train on a fixed goal-relative box x ∈ [0, 30] m from the defended goal: a keeper who has swept far upfield falls outside it and is dropped as a "sweeper rush". That is correct for normal in-goal positioning, but it makes the model hard-saturate at 30 m — it cannot place an in-possession keeper who…
Predicts where a league-average goalkeeper would position themselves given the current game state. Uses RFCDE (Random Forest Conditional Density Estimation) over HistGradientBoostingRegressor leaf assignments with weighted 2D kernel density estimation. Part of the silly-kicks soccer analytics library (GKDV research program, TF-18, Layer 2). Standard goalkeeper evaluation metrics (xGOT, save percentage, goals prevented) measure what happens after a shot is taken. Ghost-GK addresses the upstream question: given the current game state, where should the goalkeeper be standing? The model learns a league-average positional density from thousands of match frames across multiple tracking data…
Predicts where a league-average outfield rearguard defender would position themselves, per tracking frame, team, and lateral slot. The outfield sibling of Ghost-GK v1: a point estimate (ghostgrx/y, goal-relative) served as the exact pickle-free boosted HistGradientBoostingRegressor mean. Part of the silly-kicks soccer analytics library (TF-60 rest-defense; ADR-087). It is the league-average baseline the rest-defense outfield counterfactual arm differences against ("how far does the actual rest-defense shape sit from average?"). For an in-possession team's rest defense (the rearguard it keeps while attacking, to blunt the counter after a loss), the model predicts the league-average position…
This is the model card of a transformers model that has been pushed on the Hub. Users (both direct and downstream) should be made aware of the risks, biases and limitations of the model. Use the code below to get started with the model. Carbon emissions can be estimated using the Machine Learning Impact calculator presented in Lacoste et al. (2019).
Selected GLaDiT inference weights and the matching frozen UVP VGAE for initial-frame-conditioned generation of 64 future CylinderFlow frames. The source archive includes same-machine latency, Validation100 and plotting tools. 只需分享本仓库链接。DiT 权重及配套 VGAE 已公开；baseline 沿用既有已选权重和 prepared 数据，详见交接说明。 condition width512, 1000 training diffusion steps; DDIM eta0 sampling. updates and raw/EMA candidates, score0.08477181532805404 under the recorded 20-step, three independently scored sample protocol. - Parent training was paused at update811685 of the planned1000000 updates. This is the selected550k checkpoint, not the pause recovery checkpoint. - The accompanying VGAE was selected by its declared…
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.
Fine-tune of fastino/gliner2.5-base-v1 (GLiNER2, boundary architecture) for data-use mention extraction (dataset / survey / census / registry mentions in economics research papers). Labels were produced with a locally run LLM through the passage-annotation harness, with human-in-the-loop passage review and adjudication under the project doctrine. Holdout F1 is therefore agreement with the annotating agent, not owner accuracy. - NAMEDDATA — a proper name, title, or acronym of a specific data source - DESCRIPTIVEDATA — a source described in words but not named - VAGUEDATA — generic data wording with no identifiable source 2 gold mention(s) in the holdout sit inside a run the word splitter…
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…
نموذج GLM-4.6V-Flash (9B VLM، كمية Q4KS، رؤية عبر mmproj-F16) يعمل على Kaggle CPU-only عبر llama.cpp، خلف بروكسي جسر متوافق مع OpenAI، ومنشور عبر Cloudflare Tunnel باسم نطاق دائم. - API Key (حقيقي مُصادق عليه): sk-glm46v-2026-xK9mPqR3vT7wL5nB8cD2fH4j — الطلبات بدون مفتاح صحيح ترجع 401. (المفتاح القديم sk-glm46v-flash-cpu يظل مقبولًا كاحتياط.) - النفق: Cloudflare Tunnel v2 باسم glm46v-kaggle-cpu-v2 (id 345c1c71-1b97-48fd-98c7-aca9db1ecd58) — النفق القديم f1eec3dc حُذف نهائيًا مع CNAME محدّث إلى النفق الجديد. العنوان العام ثابت: https://glm46v.kenven.bond. 1. إزالة --cache-reuse 256: كان يجبر llama على prompt eval توكن-توكن (~220-430ms/توكن) → برومبت 2100 توكن ينتظر 7 دقائق قبل أول توكن.…
This is the complete TrellisMX codec-v2 checkpoint package, variant B, built on four rented NVIDIA B200 GPUs. The root contains all 42 main routed layers plus MTP45 in 172 TP4 sidecars in the existing coupled H512/H128 format. carrier/ contains the unchanged backbone, shared-expert, attention, vision, non-expert MTP, scale, configuration, and tokenizer assets from the pinned native carrier. No separate base-weight download is required. Loading requires the TrellisMX runtime; this is not a generic Transformers checkpoint. Download the complete package with hf download brandonmusic/glm-5.3-flash-tr3-fp8 --local-dir glm-5.3-flash-tr3-fp8. The root config.json is included in the complete…
Blockwise FP8 quantization of zai-org/GLM-4.7-Flash: e4m3 weights with one scale per 128×128 block and dynamic activation scaling, about half the size of the BF16 original (30 GiB vs 58 GiB). See the original model card for model details. gsm8k (5-shot, full test set) on SGLang with MTP speculative decoding: 0.809 for FP8 vs 0.819 for BF16; average accept length 2.41 vs 2.43. Attention tensor parallelism must be 1 or 2: at 4, each rank's kvbproj shard (2240 rows) is not a multiple of the 128-row block. For more GPUs, use DP attention with expert parallelism, for example --tp-size 4 --dp-size 4 --enable-dp-attention --ep-size 4 --moe-a2a-backend deepep --cuda-graph-max-bs-decode 128. MIT…
Model Collections
Hand-picked starting points, each with the reason it exists.
Collection · 4 entries
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.
Collection · 4 entries
Models that fit 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.
Collection · 6 entries
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.
Collection · 3 entries
Speech and audio models
Recognition and synthesis models, grouped so the two directions are easy to compare.
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
The same open-weight model priced by every host that serves it, per million tokens.
Research Hub
Data center trackers and maps
Moratoriums, permits, power, water and capital behind the facilities that run these models.
Method
How the Model Hub is built
Sources, evidence labels, refresh behaviour, and the limits of every comparison here.