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.
An MLX 4-bit conversion of NVIDIA's for Apple silicon. This is a community conversion. It is not made, reviewed, or endorsed by NVIDIA. For the model's design, training data, benchmarks, and intended use, see Dory 1.0 4B is a looped (latent-recurrent) normalized-GPT model trained by NVIDIA. Its layers form three blocks: input, recurrent, and output. The shared recurrent block runs several times per token, and the input block's output is added back at the start of each loop. More loops give more compute per token and better accuracy from the same weights. It mixes 1,024-token sliding-window attention with full attention, supports up to 256K tokens of context, and reasons in a block before it…
Using llama.cpp release b11259 for quantization. Don't know which to choose? Grab Q4KM (4.78GB) - usually a good mix of size and performance. Download instructions available here First, make sure you have the Hugging Face CLI installed: These quants run with llama.cpp - installable in one line via llama.app: llama-server includes a built-in chat web UI, served at http://localhost:8080 by default. These quants were made with llama.cpp release b11259 - if this model's architecture is newly supported, you'll need that release or newer to run them. They also work in: quants made using imatrix option, with a calibration corpus rendered through this model's own chat template. The corpus pairs…
This repo contains some APEX quantizations of Naphula/Goetia-26B-A4B-v1.7, created using bartowski's imatrix found here. As explained by its creators, APEX is "a novel MoE-aware mixed-precision quantization technique for llama.cpp". I suggest reading their information on the how and why of the method. Quantized using llama.cpp #1692f9e50 For roleplaying, I personally strongly recommend nohurry's SillyTavern presets for Gemma 4 models. My only contribution is compute. This is neither my merge, nor my quantization methodology. Have fun (within the constraints of the license). Google - the base model. Naphula - the Mergelord. Everyone whose finetunes were included in the merge! mudler - for…
Research-grade model with operational-style verification discipline libraryname: pytorch - weather - weather-forecasting - hurricane - tropical-cyclone - era5 - weatherbench2 - hurdat2 - geodesic-attention - physics-informed - probabilistic-forecasting - ihelix - weatherbench2/era5 - hurdat2 pipelinetag: other NatureV1 is an 89-million-parameter model of the atmosphere that forecasts global weather out to five days and, for tropical cyclones, where a storm will go, how strong it will get, whether it will intensify rapidly, and whether it will make landfall. Every number it produces comes with an uncertainty. It is built on iHELIX, an attention engine that works on the physical geometry of…
A 62.7M-parameter Parakeet-style Token-and-Duration Transducer for Arabic speech recognition — FastConformer encoder, TDT objective — trained from scratch with no pretrained encoder and no external language model. On the six test sets of the it reaches 31.34 average WER / 13.62 CER, which places it 8th of 38 evaluated systems — ahead of Voxtral-Small-24B, gemma-4-E4B, Qwen3-ASR-1.7B, nvidia-conformer-ctc-large-arabic and every Whisper model, at a fraction of their size. Every model above it is between 1B and 30B parameters, or a commercial API. Averages are recomputed over these six sets for every model so the ordering is consistent. Baseline figures are the leaderboard's published numbers.…
This gemma3 model was trained 2x faster with Unsloth
This gemma3 model was trained 2x faster with Unsloth
Core AI (.aimodel) conversion of NVIDIA's streaming speaker-diarization model (up to eight speakers, 10 ms output frames) for iOS 27+ / macOS 27+. Converted from checkpoint revision (Nemotron-3-Diarization.nemo, SHA-256 867c53f552998f772e5b5e5c082962ae85ee7ca5669c2bc17d7f615133d4e96d) with coreai-torch 0.4.1 (standard graphs) and 0.4.2 (Neural Engine encoder), coreai-core 1.0.0b2. No weights were retrained, quantized or pruned. The model is split into two graphs that a host streaming loop calls once per encoder/head (packed speaker cache + FIFO + chunk embeddings → speaker probabilities). The Arrival-Order Speaker Cache, FIFO and chunking logic of Streaming Sortformer run on the host; see…
Core ML conversion of NVIDIA's streaming speaker-diarization model (up to eight speakers, 10 ms output frames) for iOS 18+ / macOS 15+. Converted from checkpoint revision (Nemotron-3-Diarization.nemo, SHA-256 No weights were retrained, quantized or pruned. The model is split into two graphs that a host streaming loop calls once per encoder/head (packed speaker cache + FIFO + chunk embeddings → speaker probabilities). The Arrival-Order Speaker Cache, FIFO and chunking logic of Streaming Sortformer run on the host and are not part of these graphs; see "Streaming loop" below..f32le files are raw little-endian float32 in C order. The.mlpackage directories load with MLModel.compileModel(at:) /…
GGUF conversions of NVIDIA's Nemotron-3 Diarization (Streaming Sortformer v3) for transcribe.cpp, the runtime behind speaker labels in Glimpse. The model answers "who spoke when" for up to 8 speakers at 10 ms resolution. It produces no text. AMI meeting corpus, IHM test set: 16 meetings, 9.06 hours. Scored against NVIDIA's forced-alignment RTTMs with a 0 s collar, overlapped speech included, and a plain 0.5 threshold at 10 ms with no post-processing. Every row went through the same pipeline. Compared with Sortformer v2.1, errors drop by 42% and speaker confusion by 76%, and the 1-second streaming mode beats v2.1's offline mode. Throughput on an Apple M2 Pro (offline mode, times faster than…
NVIDIA's Nemotron 3.5 Bolt 24GB, repackaged so it runs entirely on the GPU on Apple Silicon. On an M5 Pro, prefill goes from 180 → 675 tok/s and decode from 12.8 → 24.6 tok/s, at the full 262K context. Weights, behavior and chat template are the same as upstream. One storage format changed. The upstream GGUF stores the input and output projections of all 40 Mamba layers (ssmin, ssmout, 80 tensors, 2.5 GB) in NVFP4. llama.cpp's Metal backend has no NVFP4 matrix-multiply kernel. Those 80 tensors silently stay on the CPU. Every token makes 80 round trips between CPU and GPU. The server sits at ~557% CPU while the GPU waits. This build converts only those 80 tensors from NVFP4 to Q80, which…
NemotronH hybrid Mamba2 + attention + MoE · 74B total / 9B active · VL (image + video) · 22 GB on disk This is NVIDIA's own Nemotron-3.5-Bolt-74B-A9B with its refusal training removed and quantized to NVIDIA's own 24GB mixed-precision layout. The multimodal vision tower is kept byte-identical to the BF16 source; it is shipped with NVIDIA's own mmproj file and the matching LiLiCorr speculative drafter. llama.cpp upstream b76b13379 or newer. Older forks cannot load Nemotron-H puzzle checkpoints. Core text + vision: With speculative decoding via LiLiCorr drafter: Thinking is a binary enablethinking chat-template kwarg (no effort levels): Default sampling: temperature=0 greedy for the cleanest…
TokenAI is a non-profit startup founded in 2025 by Assem Sabry, based in Alexandria, Egypt. The Neo model, source code, training data, documentation, and related materials are owned by TokenAI. Neo is a compact TokenAI System One decision model for selecting actions from a declared candidate set. It reads English application context and a typed decision schema, then returns calibrated outputs for choice, score, and noul in one encoder forward pass. It is not a conversational or free-form text-generation model. Training dataset: tokenaii/Neo-dataset — https://huggingface.co/datasets/tokenaii/Neo-dataset It contains the complete project source code, training code, synthetic-data generation…
weighted/imatrix quants of https://huggingface.co/prithivMLmods/NeoHorse-1-4B-Max-Heretic-Uncensored-TURBO For a convenient overview and download list, visit our model page for this model. static quants are available at https://huggingface.co/mradermacher/NeoHorse-1-4B-Max-Heretic-Uncensored-TURBO-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…
static quants of https://huggingface.co/OpenIntelligenceNet/NeoHorse-1-9-Uncensored 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…
weighted/imatrix quants of https://huggingface.co/OpenIntelligenceNet/NeoHorse-1-9-Uncensored
A 272.7M-parameter language model, pretrained from scratch on 11 Indic languages + English (Sangraha corpus), then fine-tuned on grounded multilingual QA for the Indian government-schemes / financial-banking domain (PM-KISAN, Ayushman Bharat, banking products, insurance, savings instruments, etc.). Updated in place — this repo tracks the current best domain checkpoint, not a fixed snapshot; check back for updates as fine-tuning improves. Built from custom composable primitives, structurally equivalent to Qwen3 (confirmed by direct source comparison during HF conversion) and saved in that format for standard transformers loading: tokens/phase across H100 and V100 GPUs (best validation loss…
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.