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.
Updated 2026-10-23 · How the library is built
12,579 models, sorted by most downloaded.
A variant of pfeifferj/GLM-5.3-Flash-GSQ-RCO-GGUF (3.0-bit) made for faster single-stream decoding of this 117 GB model on consumer GPUs with an expert cache. Only the non-expert Q80 weights changed (attention, shared experts, dense FFN, output head: Q80 to Q4K, ~8.3 GB of the file). All 126 routed-expert tensors, tokenembd and every F32/BF16 tensor are copied bit-for-bit from the original. Those Q80 weights are streamed by the GPU for every token, so shrinking them speeds up decoding; the experts are unchanged. Stock llama.cpp can't load GLM-5.3-Flash yet. Use neurall/llama.cpp (GLM-5.3-Flash support plus a VRAM-filling MoE expert cache). It adds real GPU acceleration: on 2 GPUs, 2x the…
A lossless NVFP4-CSF container of local-inference-lab/GLM-5.3-Flash-NVFP4 at revision 175ae8ce3b5af842b0d0140dbeb43e9cfc557c49, the GLM-5.3-Flash QAD checkpoint (NVFP4 routed experts, BF16 dense layers). CSF stores the routed-expert block-scale planes in a compressed form. Nothing else changes: FP4 weight nibbles, every other tensor, the tensor names, dtypes, shapes and the source metadata are all byte-identical. Decoding is integer arithmetic on bytes, with no requantization or fitting. The original safetensors shards can be restored bit-exactly. - Built with trellis-quant trellisquant.losslessscalecheckpoint (commit 60feca330087, family glm53nvfp4). - verification.json: every original…
This is a Docker-save runtime archive for two NVIDIA Thor T5000 nodes with driver 595.78 and SM 11.0. It contains no model weights. Download the target and DFlash2 snapshots separately from the manateelazycat model repositories. The corresponding LPK uses the selected Hugging Face / ModelScope source for both model files and this archive. Public downloads require no hub login. The historical 0.5.0 reference environment was TensorFold 0.6.1 plus Mia 1.3.2 with the validated Thor decoder and communication patches. Default capacity is 1,048,576 tokens per request, a 2,684,928-token shared KV pool, and four streams. Reference acceptance includes empty-cache CUDA JIT, vision, structured output…
The intervention source is GLM-5.3-Flash-Ablitered2 (LoRA v2). The files here are complete MLX model weights, not standalone LoRA adapters; do not apply v2 again. These releases are intended for controlled safety research and red-teaming. Reducing refusal behavior also weakens a safety boundary; read the limitations and disclaimer before use. The directories are alternative complete model formats, not parts to load together. The original block-FP8 tensors were dequantized before MLX quantization. Token embeddings and the language-model head remain BF16; affine scales and biases are FP16. Quantization can change behavior. The consuming MLX runtime must support the GLM5-Next architecture…
Mixed-precision GGUF derived from The recipe preserves higher precision in always-active paths and six boundary routed layers, with IQ2XXS as the minimum precision. Status: published. File identity and bounded single-GPU image checks are complete. General text, MTP and DGX Spark context checks remain pending. Gate/up use the same format within each layer. Each tensor uses one block format throughout; routed matrix widths remain divisible by 256. Source-weight reconstruction proxies were collected for all 126 trunk expert matrices. Q2K does not require an imatrix; optional importance weighting uses available activation statistics without changing the format or file size. Calibration covered…
This is a quantized build of GLM-5.3-Flash-Uncensored-FP8. Only the routed MoE experts are re-quantized; every other tensor is kept exactly as in the FP8 source (FP8 tensors stay FP8, BF16 tensors stay BF16). The model card below this section is the original card of the FP8 checkpoint. Total ~179 GiB (the FP8 source is 306 GiB). Format. compressed-tensors, mixed-precision with two config groups in config.json: - group0 (pack-quantized, targets re:.mlp\.experts\..): the layout llm-compressor writes, per expert projection weightpacked (int32, eight 4-bit values each), weightscale (BF16, one per 32 consecutive input-channel weights) and weightshape. - group1 (float-quantized, targets the…
A weight-space combination of two independent abliterations of the same base model, produced by different teams with different methods. The two edits turn out to lie in nearly orthogonal directions, so applying both removes the refusals that each one leaves behind — the combination is more thoroughly decensored than either parent, with no fine-tuning and no gradient steps. This card documents what was done and why it works at the weight level. It does not report harmful- capability evaluations, and standard capability benchmarks (MMLU, etc.) have not been run — see Why Q80 is the base to requantize from. Both parents are FP8 (e4m3) native. Q80 (int8 + one fp16 scale per 32 weights)…
This is a quantization of GLM-5.3-flash that is optimized to work better on older hardware. This quant was built using i-matrix GGUF quantizations with the same principle as the GLM-5.2 and GLM-5.3 e-waste editions: K-quants for the routed experts, no codebook i-quants, because on pre-AVX-512 CPUs and previous-generation datacenter GPUs (MI100 / gfx908) expert dequantization, not bandwidth, is the decode bottleneck. Format: mainline llama.cpp glm5-next (merged 2026-09-30, ggml-org/llama.cpp#27773). These files do not load in builds that only know the earlier glm5next arch from the unmerged PR #27754 (unsloth's September files are that format). Any llama.cpp from 2026-10 on loads them. The…
A lossless NVFP4-CSF container of local-inference-lab/GLM-5.3-NVFP4 at revision b472e4ee53f6a9862da5486c56c6ca21be3dab70, the GLM-5.3 NVFP4 checkpoint (744B MoE; NVFP4 routed experts, BF16 dense layers). CSF stores the routed-expert block-scale planes in a compressed form. Nothing else changes: FP4 weight nibbles, every other tensor, the tensor names, dtypes, shapes and the source metadata are all byte-identical. Decoding is integer arithmetic on bytes, with no requantization or fitting. The original safetensors shards can be restored bit-exactly. - Built with trellis-quant trellisquant.losslessscalecheckpoint (commit 6ecc90456275, family glm53744bnvfp4). - verification.json: every original…
Per-expert ARVQ v3 cold experts, NVFP4 hot experts, ARVQ-scored REAP allocation. 75/75 MoE layers replaced by the full-corpus sequential PV campaign. Other layers retain their previously published weights; see pvprogress.json. Each layer's two tensor files and reports are replaced together in one commit. Training draws sequentially from 18,001,846 text tokens at context 1024. Fixed validation and development-audit sets each contain 16,384 tokens. Adam trains FP4-constrained per-expert books and FP8-constrained per-block scales. The effective batch is 262,144 tokens, accumulated in four 65,536-token passes. Layers 4–26 use 69 updates: book/scale LR.048/.032 through update 45, then.012/.008.…
Join our WeChat and Discord community Use GLM-OCR's API GLM-OCR is a multimodal OCR model for complex document understanding, built on the GLM-V encoder–decoder architecture. It introduces Multi-Token Prediction (MTP) loss and stable full-task reinforcement learning to improve training efficiency, recognition accuracy, and generalization. The model integrates the CogViT visual encoder pre-trained on large-scale image–text data, a lightweight cross-modal connector with efficient token downsampling, and a GLM-0.5B language decoder. Combined with a two-stage pipeline of layout analysis and parallel recognition based on PP-DocLayout-V3, GLM-OCR delivers robust and high-quality OCR performance…
autotrust/GLM5.3-Flash-E192-DGX-Spark is a compact build of zai-org/GLM-5.3-Flash for desktop Blackwell systems like NVIDIA DGX Spark. It's an unofficial derivative. It keeps 192 of the 288 routed experts in each layer by Neural Architecture Search (NAS), uses NVFP4 for the experts and still activates 18 B parameters per token. The weights take 124 GiB, the smallest in the family: two DGX Sparks connected by ConnectX-7 get about 62 GiB of weights per node and the most KV room of any build, and a single 180 GB Blackwell GPU (B200/GB200) runs it with 32 GiB of KV cache. The KV cache uses FP8. The repo includes a ready-to-use NVFP4 MTP draft in mtp-nvfp4/ (6.95 GB) for speculative decoding. On…
autotrust/GLM5.3-Flash-E256-DGX-Spark is a compact build of zai-org/GLM-5.3-Flash for Blackwell systems, from DGX Spark clusters to a single B200. It's an unofficial derivative. It keeps 256 of the 288 routed experts in each layer by Neural Architecture Search (NAS), uses NVFP4 for the experts and still activates 18 B parameters per token. It's the largest architecture in the family, with the best Chinese, code and vision scores. The KV cache uses FP8. The repo includes a ready-to-use NVFP4 MTP draft in mtp-nvfp4/ (6.95 GB) for speculative decoding. On a single B200 it gives 1.84× single-stream decode speed (1.94× with 3 draft tokens), and output quality is unchanged. The model keeps the…
Core AI port of fastino/gliner2.5-multi-v1, the multilingual GLiNER2.5 boundary extractor. It is built on mDeBERTa-v3-base and has 287M parameters. It keeps the upstream decision behavior. All credit for the model goes to Fastino and the GLiNER2 authors. The original model card is included as - Checkpoint fastino/gliner2.5-multi-v1 at revision 12fc40399dae672ce840c5e3c50a92340bff3c8c. 55656fb (for preprocessing and decoding). - coreai-torch 0.4.3, coreai-core 1.0.0b3, Torch 2.11.0. - Learned weights are unchanged (FP16). The export only rewrites graph structure: - Constant DeBERTa relative-position buckets. The float log-bucket path produces wrong indices inside converted graphs. - A…
公开交接包：V7b11 长窗口 tokenizer + realcam64 因果相机 DiT。仓库根不是工作目录，而是 tar 分卷（约 60GB，不压缩）。单文件超过 Hugging Face 上限，所以拆成多个 GPSToken-handoff.tar.part-。 拼完后先读 README.md 和.cursor/skills/gpstoken-migrate/SKILL.md。续训 DiT：export WANDBAPIKEY=... 然后 bash ops/resumedit.sh。本包不含任何密钥。 分卷校验和在 GPSToken-handoff.tar.sha256。
Five independently trained, model-only checkpoints at 30,000 steps, all from nvidia/GR00T-N1.7-3B. Every folder named checkpoint-30000 is directly loadable with its own config, processor, and statistics. All use observation history 2, action horizon 10, and global batch size 32. PickCube V4 force and MoE use the invalid-action loss mask and recovery sample weights. The files are the original saved checkpoints, including training configuration. manifest.json lists SHA-256 checksums for every file. Training completion and checkpoint validation do not imply an evaluation score.
압축 가능성 라벨로 게이트를 학습한 액션 양자화(ATQ) 체크포인트다. 하나의 정책이 미세(1x) 디코더와 압축 디코더를 함께 갖고, VLM 라벨에서 배운 conf 가 둘 중 어느 군을 쓸지 고른다. 라우터는 그 군 안에서 horizon 만 고른다. (VLM 전용, 접촉 열 없음 · 16,286행 · 1,693에피 · stride 16) moeexperthorizons = [16, 9, 5, 8] · confthreshold(tau) = 0.55 · discreteactiondims = [6] (그리퍼는 절대 명령이라 · actionmergereduction = sum 회전 병합은 SO(3) 다(rotationmergespec 이 config 에 있다). 압축 블록의 회전 다시 정규화한다. scipy 대조 각도 오차 1e-14도. LIBERO 는 5 fine 스텝마다 재계획한다. 압축 행 하나는 fine 액션 2~3개의 합이므로 같은 배속이 되고, 배속을 움직이는 손잡이는 conf 게이트 하나다: 넘으므로 OSC 팔 컨트롤러의 입력 클립을 제거한 조건에서 평가했다(그리퍼 그대로). 게이트는 벤치마크가 실제로 깨지는 순서를 따른다 -- 압축에 강한 liberoobject 를 가장 많이 압축하고, 2배에서 -0.160 으로 무너지는 liberospatial 은 거의 압축하지 않는다. 브랜치 jimin-dev-label-gated.…
압축 가능성 라벨로 게이트를 학습한 액션 양자화(ATQ) 체크포인트다. 하나의 정책이 미세(1x) 디코더와 압축 디코더를 함께 갖고, VLM 라벨에서 배운 conf 가 둘 중 어느 군을 쓸지 고른다. 라우터는 그 군 안에서 horizon 만 고른다. (VLM 전용, 접촉 열 없음 · 16,286행 · 1,693에피 · stride 16) moeexperthorizons = [16, 7, 3, 8] · confthreshold(tau) = 0.55 · discreteactiondims = [6] (그리퍼는 절대 명령이라 · actionmergereduction = sum 회전 병합은 SO(3) 다(rotationmergespec 이 config 에 있다). 압축 블록의 회전 다시 정규화한다. scipy 대조 각도 오차 1e-14도. LIBERO 는 5 fine 스텝마다 재계획한다. 압축 행 하나는 fine 액션 2~3개의 합이므로 같은 배속이 되고, 배속을 움직이는 손잡이는 conf 게이트 하나다: 넘으므로 OSC 팔 컨트롤러의 입력 클립을 제거한 조건에서 평가했다(그리퍼 그대로). 게이트는 벤치마크가 실제로 깨지는 순서를 따른다 -- 압축에 강한 liberoobject 를 가장 많이 압축하고, 2배에서 -0.160 으로 무너지는 liberospatial 은 거의 압축하지 않는다. 브랜치 jimin-dev-label-gated.…
This repository contains ten independent, model-only checkpoint-30000 directories. Each was trained for 30,000 steps with two image and joint-state observations, action horizon 10, and global batch size 32. The original checkpoint configuration, processor, normalization statistics, tokenizer files and safetensors are retained in each directory. These training runs did not include evaluation; no success rate is claimed. The eight force runs use a four-frame six-axis wrench history, an MLP force encoder, and one force condition token. PickCube V4 MoE uses a six-frame wrench history, a temporal Transformer force encoder, one force token, four experts, and top-1 routing. The two PickCube V4…
This repository contains the model-only checkpoint saved at 30,000 training steps. The files at the repository root are the original contents of checkpoint-30000. Full model and processor settings are in config.json, processorconfig.json, and experimentcfg/. The training run did not include an evaluation. Source checkpoint: outputs/moeobs2h10fourtasks20260926191812/pair67/usb/moe/checkpoint-30000.
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.