28 layers, width 1536, 16 heads, SwiGLU FFN 4096, RoPE, RMSNorm, tied embeddings, context 256. IMPORTANT: Fresh weights trained from scratch on English FineWeb-Edu (language modeling) and Smol-SmolTalk (chat). This model is experimental. A public checkpoint is not a guarantee of fluent answers. Consult trainingstate.json to check its training phase and token counts. Tokenizer: new ByteLevel BPE with 16,000 IDs. Chat prompts use question and responses end in. Training sources: https://huggingface.co/datasets/HuggingFaceFW/fineweb-edu and https://huggingface.co/datasets/HuggingFaceTB/smol-smoltalk. Respect upstream dataset terms. Training uses a single Kaggle GPU and restarts across sessions.…
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.
12,579Models
3,583Datasets
291Papers
7,578Publishers
12,850Sourced relationships
Updated 2026-10-23 · How the library is built
12,579 models, sorted by most downloaded.
Overlay metadata repo for using FastVideo/FastVideo-FastH3-8-Step-V2 with native SGLang Diffusion MiniMax-H3 support. This repo contains only overlay metadata and materialization logic. Source weights remain in FastVideo/FastVideo-FastH3-8-Step-V2; SGLang's registry pins a revision of this repo so re-materializing the model does not drift. The materializer reshapes the flat native-Diffusers release into the base MiniMax-H3 partition layout SGLang's loaders consume: - modelindex.json declares the t2va-only release contract (tasks: ["t2va"], video/audio sigma shifts 10/3, and the trained eight DMD rungs [999, 874, 749, 624, 500, 375, 250, 125]). - overlay/materialize.py re-serializes the…
The FastH3 8-Step V2 checkpoint from FastVideo. It generates synchronized video and audio from text with eight transformer forwards. This step-1300 model was trained with data-free DMD2 and VSA-H3 at 80% sparsity. Install uv, then use the CUDA 13 / Blackwell path below. It selects FastVideo's published CUDA kernel wheel instead of compiling the kernel locally. See the for other platforms. The tested defaults use four B200 GPUs and the trained eight-forward schedule. On other multi-GPU CUDA systems, follow the installation guide and add --no-replicated-dit --vsa-kernel triton --no-fa4. The GPU count must divide H3's 56 attention heads. This checkpoint supports text-to-audio-video generation.…
本仓保存 RoboTwin 实测模型、TDAA codec、原始结果和 4,000 个视频。 在 FastWAM 中接入预训练 TDAA version3bin24 编解码器，将 [32,14] 动作块编码为 [8,16] latent。策略在 latent 空间做 flow matching，再解码为 32 步绝对关节动作。decoder 使用任务向量和由已执行动作历史的 DCT24 特征生成的 phase；每个 episode 重置历史。 frozen 控制整个 TDAA codec，FastWAM 策略仍参与训练。联合训练额外加入动作重构与进度预测损失。动作 token 从 32 个变为 8 个，仅表示动作表示压缩；本次没有端到端加速测量。 - TDAA codec 与配置：codec.pt、config.json、metadata.json、datasetstatistics.json、taskembeddings.json，来自 80,000 步 AE。 代码仓的下载工具按固定版本获取文件并校验 SHA-256。在代码仓安装环境后运行： 权重保存到 checkpoints/released/{official,tdaafrozen}/step002725.pt。Wan VAE/T5/tokenizer 等基础模型、训练数据和 RoboTwin 仿真资产需另外准备，详见 GitHub 复现说明。本次不包含 optimizer/scheduler 完整训练状态。 以下为 FastWAM 基线与 TDAA Frozen=True 两组实测训练的共同设置，已与各自保存的…
A Haitian Creole (Kreyòl) text-to-speech model: Kyutai's Pocket TTS (French 24-layer model) fine-tuned on Kreyòl speech. It runs on a CPU and streams: the first audio arrives in about 0.12 s on a laptop. It is the voice of Klara, a Kreyòl voice assistant that runs offline. 80 held-out Kreyòl sentences (none in training), read aloud and transcribed back by two speech recognizers; character error rate, lower is better. The French model already reads Kreyòl at about 20% CER (its errors are French spelling habits, silent final letters above all); the fine-tune keeps its tokenizer and text embedding and teaches it Kreyòl spelling. Nine Kreyòl voices come with the model, in voices/. Each was made…
This is the Haitian Creole fine-tune of Pocket TTS (from french24l), exported to ONNX for on-device use with ONNX Runtime. It is what the Klara mobile app (Expo / React Native, Android) loads. The stock export passes a fixed-size KV cache (100 MB) in and out of the graph every frame. On a phone, that copy cost more than the model itself. This export takes each layer's past keys and values at their used length, and returns only the new positions: The caller appends new to its own buffers. The Klara app uses one time-major buffer per layer and feeds zero-copy prefix views. 1. Voice: run once with textembeddings = [bosbeforevoice, voice frames] and empty past. Keep the outputs as that voice's…
weighted/imatrix quants of https://huggingface.co/Vortex5/Fenrir-X-26B-A4B For a convenient overview and download list, visit our model page for this model. static quants are available at https://huggingface.co/mradermacher/Fenrir-X-26B-A4B-GGUF This is a vision model - mmproj files (if any) will be in the static repository. 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…
Meridian.AI is a finance-specialized language model that continuously fine-tunes a Qwen2.5-0.5B backbone every hour on 25+ finance and math datasets, using Elastic Weight Consolidation (EWC) to prevent catastrophic forgetting across training sessions. The entire pipeline runs unattended on free GitHub Actions infrastructure — no GPUs. The deployed checkpoint is a standard Qwen2 model — trustremotecode=True is not required. Inputs are formatted with the ### Instruction: / ### Response: template used during training. A weighted streaming mix of 25+ finance and instruction datasets, including gbharti/finance-alpaca, sujet-ai/Sujet-Finance-Instruct-177k, nvidia/OpenMathInstruct-2…
static quants of https://huggingface.co/whcl412/FinchAI-Rune1-Preview-f16 For a convenient overview and download list, visit our model page for this model. weighted/imatrix quants seem not to be available (by me) at this time. If they do not show up a week or so after the static ones, I have probably not planned for them. Feel free to request them by opening a Community Discussion. 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…
Firebird-ModernBERT-512-RW is an experimental post-trained variant of It is a ~149M parameter ModernBERT binary classifier for distinguishing: - 0 — HUMAN The maximum sequence length is 512 tokens. This checkpoint was produced through reward-weighted classifier post-training. The original Firebird checkpoint was kept frozen as a reference model. Training examples were scored by the original classifier, difficult examples received larger loss weights, and the post-trained model was constrained against the frozen reference using a KL penalty. L = weightedcrossentropy + beta KL(reference || policy) Hard human examples received greater weighting than ordinary examples because one goal of the…
Fleet is a dynamically extensible Semantic Inference Language Model (SiLM) built on Qwen/Qwen3-4B-Instruct-2507. Fleet is not an external prompt router. The routing system lives inside the causal language model itself through the custom FleetForCausalLM architecture. Fleet replaces selected Qwen projections with native FleetLinear modules. The model also contains a dynamic semantic prototype bank: Routing therefore happens inside model inference, rather than through an external classifier that chooses a different model before generation. This checkpoint currently contains six specialists: - mathreasoning - medicalreasoning - legalops - intent - functioncalling - structured The base Qwen…
An archived, matched pair of RG-Diff and SC-AR adapters for FlexDLLM: Diffuse When Possible, Autoregress When Necessary. The two stages run on a shared LLaDA-8B-Instruct backbone. RG-Diff emits during diffusion; SC-AR generates a short autoregressive repair, and diffusion resumes. This is a research preview of the existing math-oriented checkpoint pair. The release has passed a real-model ASK/repair/resume smoke test. The paper's full benchmark scores have not been reproduced for this bundle. The download contains approximately 521.5 MiB of adapters, tokenizer and configuration. The 8B backbone is downloaded separately by the loader. Both adapters use LoRA rank 32, alpha 64. RG-Diff is…
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.