Image and Text to Text Models

This is an uncensored version of deepseek-ai/DeepSeek-V4-Flash-Vision-Exp 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 11-28 are being ablated (1-based indexing) and all expert modules were not ablated. 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 outputs. - Support our work on Ko-fi!

Model · Image and text to text

Huihui-GLM-5.3-Flash-abliterated-FP4-Local

Edp1096

Unofficial NVIDIA FP4-Local derivative with Huihui Q8 deltas applied. Applied delta: DQ(Huihui Q8) - DQ(Unsloth Q8). The conversion used these exact revisions. - Updated 44 attention-output/shared-expert-down matrices in layers 15–36 and 22 scale tensors - Preserved all other tensors, including routed experts, CSF scales, MTP, vision, and tokenizer files - Requantized in the existing FP8 formats; no additional QAD training or activation calibration Requires two DGX Sparks (TP2) and B12X vLLM with CSF support. See runtime-configuration.json for exact tested settings. Cold-start loading caused memory pressure and some host swapping. - All 42 weight files verified; 8/8 smoke checks passed…

This is an uncensored version of zai-org/GLM-5.3-Flash 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. GGUFs come from unsloth/GLM-5.3-Flash-GGUF. Only layers 15 to 35 (0-based indexing) have been ablated, while the other layers and all expert modules remain unablated. To run, please unslothai llama.cpp PR, - 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 outputs.…

This is an uncensored version of zai-org/GLM-5.3-Flash 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. The latest version of llama.cpp now officially supports zai-org/GLM-5.3-Flash. We have updated the first file of each quantized version, so you only need to download the first file. GGUFs come from unsloth/GLM-5.3-Flash-GGUF. Only layers 15 to 35 (0-based indexing) have been ablated, while the other layers and all expert modules remain unablated. To run, please unslothai llama.cpp PR, - Risk of Sensitive or Controversial Outputs: This model’s safety filtering has…

An FP8 quantization of huihui-ai/Huihui-Qwen3.8-27B-abliterated that reproduces Qwen's official FP8 recipe byte for byte on every tensor the abliteration did not touch. The weights are block-wise FP8 E4M3 with 128×128 scales, in the same layout as Qwen/Qwen3.8-27B-FP8. The MTP head and vision tower are preserved, and the checkpoint is 28.75 GiB. It loads through the same code path as the official FP8 checkpoint. This model was built and validated as part of a single-GPU deployment study. The full notes, scripts and raw data are at github.com/gyang274/ms-Qwen3.8-27B. The quantizer (quantizefp8like.py) uses the official FP8 checkpoint as a layout template. It quantizes exactly the tensors…

This is an uncensored version of Qwen/Qwen3.8-27B 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. The newly added Huihui-Qwen3.8-27B-abliterated-Ternary series come from prism-ml/Ternary-Bonsai-2-27B-gguf have been ablated, while the other layers remain unablated. It may come with a small disclaimer warning. The size after conversion may differ from the original GGUF (Some of the weights are converted from PTQ1 to Q2K or Q3K.). This is just a test/validation. The ternary hybrid-attention kernels live in the PrismML-Eng/llama.cpp fork.…

This is an uncensored version of Qwen/Qwen3.8-27B 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. The newly added Huihui-Qwen3.8-27B-abliterated-Underdog-Saluki series come from ConwayResearch/Underdog-Saluki-27B-1.0 and ukisai/Swift-1.5-Qwen3.8-27B-GSQ-RCO-GGUF. Only layers 22 to 52 (0-based indexing) have been ablated, while the other layers remain unablated. It may come with a small disclaimer warning. Due to ConwayResearch/Underdog-Saluki-27B-1.0 not providing an imatrix file, we will attempt to perform ablation by mixing the GGUF…

This is an uncensored version of Qwen/Qwen3.8-Flash-Next 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. This version Compared to the first version, the ablation emphasis has been adjusted, and none of the experts have been ablated. Suitable for novel writing and programming. The newly added Swift series come from ukisai/Swift-1.5-Qwen3.8-Flash-Next-GGUF. The unsloth series come from unsloth/Qwen3.8-Flash-Next-GGUF. This unsloth and Swift series supports the latest versions of both Strata and llama.cpp. Strata has byte-level and SHA256…

Model · Image and text to text

HyzeK3

Hitesh Vinothkumar

Kimi K3 is an open-weight, native multimodal agentic model and our most capable model to date. It is a 2.8T-parameter model built on Kimi Delta Attention (KDA) and Attention Residuals (AttnRes), with native vision capabilities and a 1-million-token context window. It is the world's first open 3T-class model, designed for frontier intelligence across long-horizon coding, knowledge work, and reasoning. All Kimi K3 results are obtained with reasoning effort set to 'max' and temperature = 1.0. For single-step tasks, such as GPQA Diamond, HLE-Full, and vision benchmarks without tools, we set top-p = 0.95; for agentic tasks, we set top-p = 1.0. For HLE-Full, MMMU-Pro, CharXiv (RQ), MathVision…

A multi-image chain-of-thought fine-tune of OpenGVLab/InternVL3-8B, merged into standalone weights so it loads with trustremotecode=True in transformers or vLLM without the adapter. final model. - LoRA, r=64, alpha=128, dropout=0.05, on q/k/v/oproj and gate/up/downproj - modulestosave: ["visionmodel", "mlp1"] — the vision tower and the projector were trained in full rather than low-rank adapted, which is why the adapter was 1.3 GB and why merging is required for serving Accuracy against the base model. The prompt is stated per row, because it matters more than the fine-tune does: rows are only comparable to other rows evaluated the same way. MIRB and MuirBench/BLINK use different image…

Model · Image and text to text

Inverse-LLaVA-7B

Xuhui Zhan

Inverse-LLaVA maps language features into the visual feature space and fuses them inside the language decoder. This is the full-data 7B checkpoint trained jointly on visual instructions, without a separate alignment-pretraining stage. On Linux with a CUDA-compatible PyTorch installation, install the source package: Use a commit hash in revision= to pin a particular Hub snapshot. The weights file has the stable identity below; the loader verifies its checksums before constructing the model. It downloads the pinned Vicuna and CLIP weights separately when they are not cached. The 695 MB delta is not the total download or GPU-memory requirement. Allow space for the approximately 14 GB language…

static quants of https://huggingface.co/irwanworld/Irwan-LightOnOCR-3-0.8B-abliterated 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…

4-bit AWQ quantization of autotrust/JEV-27B-VL, built with llm-compressor and served with vLLM. System 1 (POST /v1/decide) runs on a single 24 GB GPU — the fp8 build needs two — and on every axis we measured it is faster than the two-GPU fp8 deployment it replaces. The System 1 adapter is merged into the weights (including the decision head), so no --enable-lora is needed. The decision head (lmhead) and the vision tower are kept in bf16; every other text layer is quantized, including the Gated DeltaNet projections of the 48 linear-attention layers — most Qwen3.5-family quants exclude those and therefore do not fit on one 24 GB card. The AWQ step matters: a data-free RTN build of the same…

Model · Image and text to text

Jev-Omni-MLX-4bit

Leo Zhang

An Apple Silicon MLX conversion of akhilaaa3/Jev-Omni for local inference on a Mac mini with 16GB unified memory. This repository keeps the upstream Jev-Omni unified multimodal checkpoint and its trained 256-way decision head, then converts the language-model weights to 4-bit affine quantization with group size 64. The vision weights remain BF16 and the decision head remains FP32. It is an independent conversion. It is not an official TypeSafe Jev release and does not claim to reproduce TypeSafe's proprietary system. It is also not a new fine-tune. - Mac mini, Apple M4, 10 CPU cores, 16GB unified memory - macOS 26.5.1 - Python 3.13.12 - MLX 0.32.2 - MLX-VLM 0.7.1 - single request, batch…

This model was converted to MLX format from Qwen/Qwen3.5-2B using mlx-vlm version 0.3.12. Refer to the original model card for more details on the model.

A bilingual vision-language model built on Keural Mid, a 183.6M-parameter vision encoder trained from random initialization — not a CLIP, ViT or DINO derivative. This repository contains the trained bridge only: a projector and LoRA adapters, 37 MB total. The encoder and decoder load from their own repositories. A vision-language model can post respectable scores while barely using the image — "what colour is the banana?" is answerable from language priors alone. Every benchmark below was therefore run twice: once normally, once with the visual tokens removed and everything else held constant. The delta is the encoder's contribution. It is the one number a language model cannot manufacture.…

Kimi K3 is an open-weight, native multimodal agentic model and our most capable model to date. It is a 2.8T-parameter model built on Kimi Delta Attention (KDA) and Attention Residuals (AttnRes), with native vision capabilities and a 1-million-token context window. It is the world's first open 3T-class model, designed for frontier intelligence across long-horizon coding, knowledge work, and reasoning. All Kimi K3 results are obtained with reasoning effort set to 'max' and temperature = 1.0. For single-step tasks, such as GPQA Diamond, HLE-Full, and vision benchmarks without tools, we set top-p = 0.95; for agentic tasks, we set top-p = 1.0. For HLE-Full, MMMU-Pro, CharXiv (RQ), MathVision…

Kimi K3 on a single NVIDIA A100 80GB. A weight-only quantisation of Moonshot AI's Kimi K3 (2.8T total / 104B activated parameters) that loads and generates on one A100 80GB GPU, with the routed experts held in host RAM. No Ampere-targeted K3 build existed for vLLM, so this was made to fix that gap. Routed experts and attention re-encoded from MXFP4/BF16 into compressed-tensors pack-quantized, served by vLLM's Marlin kernels. Activations stay BF16 (W4A16 / W8A16). Round-to-nearest only — no calibration data, so no dataset is baked into these weights. Errors were measured by round-tripping each tensor through compressed-tensors' compress()/decompress(). Weight error is a proxy, not a quality…

No model weights or adapters are available to download. This repository contains a model card only. It is not a loadable Transformers model. No release date is promised, and no third-party weights have been uploaded. KnowinRouter jointly selects one capability expert and optional knowledge/perception under a bounded acquisition budget. The intended portfolio uses a shared Qwen3-VL-8B backbone and three independent LoRA adapters: General (continued pretraining), Spatial Reasoning, and Spatiotemporal. A frozen Qwen3-VL-2B descriptor feeds the utility router. After one acquisition round, the Packet Verifier retains valid evidence with positive predicted gain while keeping the expert fixed. The…

Model · Image and text to text

Legend_ocr_qwen2-highavg

Jackson Kahungu

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).

Model · Image and text to text

Legend_ocr_qwen3-highavg

Jackson Kahungu

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).

Model · Image and text to text

Legend_ocr_rolm-highavg

Jackson Kahungu

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).

Model · Image and text to text

LightOnOCR-3-0.8B-ONNX-web

superPDF

lightonai/LightOnOCR-3-0.8B by LightOn, converted to ONNX for superPDF, where it powers high-accuracy reading in the browser (WebGPU, superpdf.ai) and the command-line tool (default model). It runs with Transformers.js on ONNX Runtime (WebGPU), in the browser and in Node. from the unused multi-token-prediction layer). Every weight in those graphs is recomputed from LightOn's weights and written at the same place, so the graphs themselves are unchanged. asymmetric with zero points, computed with ONNX Runtime's own quantizematmul4bits); vision encoder float16. Norm weights are stored as 1 + w and the gated delta net's A as -exp(Alog), as the builder expects. - These are the same bytes as…

Model · Image and text to text

LightOnOCR-3-4B-ONNX

superPDF

lightonai/LightOnOCR-3-4B by LightOn, converted to ONNX for superPDF, where it powers high-accuracy reading in the command-line tool and MCP server (--model 4b, model: "4b"). It runs with Transformers.js on ONNX Runtime (WebGPU), in the browser and in Node. from the unused multi-token-prediction layer). Every weight in those graphs is recomputed from LightOn's weights and written at the same place, so the graphs themselves are unchanged. asymmetric with zero points, computed with ONNX Runtime's own quantizematmul4bits); vision encoder float16. Norm weights are stored as 1 + w and the gated delta net's A as -exp(Alog), as the builder expects. - Converted by superPDF with the converter below.…

Model · Image and text to text

linne-oak-1-0

Hert4

linne-oak-1.0 is a 31B decision model. Give it a state (text, a document, JSON, a conversation or an agent trajectory) and a question with a set of options, and it returns one option together with a probability for every option, in a single forward pass. It is the larger sibling of linne-fern-1.0: linne-fern is a 27B causal model that also chats, linne-oak is a dedicated decision model. - Decision Index public score 63.94, measured with the official kit (edition 0.3, all 37 public benchmarks, 140,620 requests, complete). - Reads the whole prompt at once: the full-attention layers attend in both directions over the prompt (sliding-window layers keep their native behavior), so the decision…

Model · Image and text to text

Lite-OmniJev

Wang

冻结 Gemma 4 E4B 基座 + 决策 LoRA，每题一次前向读出全部选项概率。 支持 Choice（2–20项）、Score（2–10级）、yes/no，覆盖文本、图像、截图、音频与视频。 API直接返回答案概率、置信度和结构化判断。 MMAU音频路由基座方案采用65.9%的实测结果；默认部署脚本采用全模态adapter。 真实模型工作台通过三类判断、图像/截图/音频/视频、cURL重放及错误恢复验证。 GGUF三档在20个公开文本案例上均与参考选答20/20一致，截图、音频和视频运行探针全部成功。 环境：Linux x8664、Python3.11、兼容CUDA13的NVIDIA驱动；FP32配方在H20 96GB实测显存约39GB。 终端B启动API和工作台： 打开本地工作台，或调用API： 每题返回probabilities、confidence及相应的choice、score或noul。 Score为从0开始的期望等级；Noul为yes的概率。部署脚本使用原始答案概率（calibration: unfitted）。 媒体通过服务端受信文件路径和SHA256引用： type也支持audio/video；视频用videoframes选择8或16帧。 GGUF截图采用RGB、长边≤1344；完整命令见gateway/GGUF.md。

LLaVA is an open-source chatbot trained by fine-tuning LLaMA/Vicuna on GPT-generated multimodal instruction-following data. It is an auto-regressive language model, based on the transformer architecture. LLaVA-v1.5-7B was trained in September 2023. Paper or resources for more information: https://llava-vl.github.io/ Llama 2 is licensed under the LLAMA 2 Community License, Copyright (c) Meta Platforms, Inc. All Rights Reserved. Where to send questions or comments about the model: https://github.com/haotian-liu/LLaVA/issues The primary use of LLaVA is research on large multimodal models and chatbots. The primary intended users of the model are researchers and hobbyists in computer vision…

Model · Image and text to text

Lodestar-4B

StartLux

Lodestar-4B is a 4-billion-parameter decision model. You give it a state (plain text, JSON or a long document) and one or more typed questions; it returns a probability for every listed option. Each answer is read from a single forward pass. The model never generates free text, so there is nothing to parse and no output length to budget for. (Probabilities rounded to three decimals.) The same call from Python, run inside the downloaded folder: A score question takes its levels as a list, lowest first, and returns {"type": "score", "score":, "probabilities": {"0": p0, "1": p1,...}}. Every question is rendered into one chat prompt (thinking disabled): The probability of each option is the…

a page (a book, a paper, a report) and it returns structured HTML: headings, paragraphs, tables and formulas marked separately, flowcharts as Mermaid, sheet music as ABC notation. It runs entirely on your own machine. The original GGUF lives under a single personal account, on a single host. If it is renamed, deleted or re-uploaded, people who depend on it lose their download, and access to Hugging Face from mainland China is unreliable. So we keep the original files, unmodified, on ModelScope and on our own Hugging Face account. File names, sizes and sha256 digests all match the original. Reproducible: we also converted and quantized the official Alibaba weights ourselves. The quantized…

static quants of https://huggingface.co/cagataydev/loopl-0.8b For a convenient overview and download list, visit our model page for this model. weighted/imatrix quants are available at https://huggingface.co/mradermacher/loopl-0.8b-i1-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: https://gist.github.com/Artefact2/b5f810600771265fc1e39442288e8ec9 See…

weighted/imatrix quants of https://huggingface.co/cagataydev/loopl-0.8b For a convenient overview and download list, visit our model page for this model. static quants are available at https://huggingface.co/mradermacher/loopl-0.8b-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…

static quants of https://huggingface.co/cagataydev/loopl-2b 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 quant And…

static quants of https://huggingface.co/cagataydev/loopl-vl-2b 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 quant And…

LT-OPD trains vision-language models with fewer visual tokens through on-policy self-distillation. This Qwen3.5-4B model combines CDPruner with an MLP that aggregates discarded tokens into their nearest retained tokens, preserving their original positions without adding tokens. Use Python 3.12 and the CUDA environment in the installation guide. Load the model with LT-OPD's visual-token compression runtime

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).…

Model · Image and text to text

Mephisto-1.5-4B

Model Merge JP

Mephisto-1.5-4B is a 4B-parameter agentic language model constructed via multi-stage model merging using mergekit. The model integrates reasoning, coding, and agentic capabilities into a single 4B checkpoint based on the Qwen3.5-4B architecture, targeting performance exceeding Mephisto-4B-v2 on Japanese and agentic benchmarks while maintaining deployment feasibility on consumer hardware (RTX 3060 12GB). The model is constructed through a three-stage merging pipeline: Normalized spherical linear interpolation (NuSLERP) on task vectors derived from a common base model (Qwen3.5-4B). This stage fuses a reasoning-distilled model (Opus-reasoning distillation) with a code-specialized model.…

Model · Image and text to text

meta-models.Muse-Spark-Open-816B-A42B-GGUF

DevQuasar

'Make knowledge free for everyone'

Model · Image and text to text

MIMO-V2.6-DERISKED-MXFP4

Sir Frosty

Behaviorally modified MXFP4 research checkpoint derived from XiaomiMiMo/MiMo-V2.6-Flash-RL. DERISKED identifies the Blackfrost release family. It is not a claim of zero refusals, complete safety, harmlessness, or production readiness. Refusal measurements will be published only after evaluation of the exact uploaded artifact is complete. MIMO-V2.6-DERISKED-MXFP4 is an independent Blackfrost research derivative of XiaomiMiMo/MiMo-V2.6-Flash-RL. The checkpoint retains the upstream MiMo V2.6 architecture, tokenizer, multimodal components, native chat template, and speculative-decoding assets while preserving its mixed-precision MXFP4 deployment layout. The behavioral modification is encoded in…

Model · Image and text to text

MiMo-V2.6-Flash-RL-Mixed-Quant-GGUF

Sunghoon Baek

I work on making large language models practical on constrained hardware through mixed quantization, inference optimization, and serving experiments. Contributions help cover calibration, GPU compute, storage, and testing so these results can be published openly. This is a mixed-precision conversion of XiaomiMiMo/MiMo-V2.6-Flash-RL, pinned to revision 3b38d063180c3e4aed9691fdc735f3d10b266ee4. The source stores routed expert weights as packed MXFP4. Two logical weights occupy each packed byte, so counting stored tensor elements can produce a roughly 159B figure. The expanded language trunk contains approximately 308.78B logical parameters; the root checkpoint including embedded MTP and media…

Model · Image and text to text

MiniCPM-V-4.7-1B-GGUF

Nguyen Quoc Vinh

Community GGUF quantizations of openbmb/MiniCPM-V-4.7-1B. If this GGUF made your day easier, a coffee would make mine. I build and test these releases myself. Your coffee helps keep me going. Thank you for supporting this work. MiniCPM-V 4.7 1B is OpenBMB's compact multimodal model for image and video understanding. The upstream model describes single-image, multi-image, and video use, with an efficient vision encoder and on-device deployment targets. See the official model card and the upstream deployment guide. The upstream configuration declares a 262,144-token text context. The example below uses a 4,096-token context; long-context behavior has not been evaluated in this release. This…

Model · Image and text to text

Muse-Glimmer-30B

Ares Realm Studios

Muse Glimmer is a 30-billion-parameter causal language model with a dedicated perception encoder, distilled from Muse Spark and purpose-built for autonomous agentic tasks on consumer hardware. The model integrates multi-step reasoning, reliable tool use, multimodal understanding, and failure recovery into a single model that runs locally without requiring cloud infrastructure or network access. Building effective agents requires key capabilities working together to achieve the user’s goals. Muse Glimmer is trained and evaluated on these capabilities: End-to-end Agentic Task Completion. Muse Glimmer achieves strong success rates on full-task benchmarks including DeepSearch QA, MCP-Atlas…

Four checkpoints, not one model. This repository is a controlled experiment: two training arms × two seeds, 10,000 steps each, identical in every respect except how packed training rows are masked. The comparison is the artifact. Any single checkpoint here is a weak 230M research model and is not meant for use. Each arm lives on its own branch: Weights are fp32, exactly as trained — these are the tensors every number below was computed from. Optimizer state is not included. nanoVLM packs several unrelated samples into one 4096-token row to avoid padding waste. Upstream's packing applies a single causal mask over the whole row, so a sample attends into the earlier, unrelated samples packed…

Model · Image and text to text

Nemotron-3.5-Bolt-74B-A9B-24GB-Metal-GGUF

Uraion Labs

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…

Join our WeChat or Discord community. Check out the GLM-5.3-Flash blog and GLM-5 Technical report. Use GLM-5.3-Flash API services on Z.ai API Platform. We introduce GLM-5.3-Flash, the first natively multimodal model in the GLM-5 series. With 320B total parameters and just 18B active parameters, it outperforms GLM-5.2 across benchmarks and real-world workloads at one-tenth the price, while approaching Claude Opus 4.8 on coding and agentic benchmarks. GLM-5.3-Flash starts from a newly trained base model, with its architecture and training recipe redesigned around capability and efficiency. For the first time in the GLM series, we introduce a hybrid architecture combining sparse and linear…

Native MLX port of NVIDIA NV-Reason-CT for Apple Silicon. Run 3D CT reasoning and report generation locally on Mac. - Local NIfTI input with Apple Silicon GPU acceleration - FP32 inference using MLX; no PyTorch or CUDA required MLX port by Joseph Sandoval. Requires the companion nv-reason-ct-mlx runtime. Generic mlx-vlm and mlx-lm examples do not support this 3D CT model. Apple Silicon Mac, macOS, Python 3.12 and uv. Weights occupy 17.4 GB. Measured peak MLX memory use is about 22.4 GB; allow additional unified memory for preprocessing, macOS and other apps. To ask a question about the CT, set --prompt to your question. Input is one 3D CT volume in NIfTI format (.nii or.nii.gz) with…

Model · Image and text to text

NVIDIA-Nemotron-3.5-Bolt-74B-A9B-Abliterated-24GB-GGUF

Josh Bolding

An uncensored edit of NVIDIA Nemotron-3.5-Bolt-74B-A9B, a hybrid Mamba/MoE model, produced by editing NVIDIA's shipped 24GB 2-bit GGUF (nvidia/NVIDIA-Nemotron-3.5-Bolt-74B-A9B-24GB-GGUF) directly — no BF16 requantize at any point. That matters for this model specifically. Bolt is a LatentMoE: every routed expert sums through one dense ffnlatentup per layer, and the experts themselves are ~17GB of IQ2XXS weights produced by NVIDIA's quantization-aware distillation. A conventional abliterate-then-requantize pipeline would throw that QAD work away. Instead the projection is applied in place and codes are re-picked under the original block scales, so at λ=0 the output file is byte-identical to…

One street photo in, a structured Japanese observation record out. OddEye is the street-reading model of michiyomi. It is built on Qwen3.8-27B and compressed to 4-bit (NVFP4, about 20 GB), so a single GPU — even an RTX 5090 — can run it. It keeps what it sees (observations) and what it infers (interpretations, each with a confidence label) in separate fields. The record has 13 top-level fields, including objects, text on signs, relations, surfaces, standard road facets (lanes, sidewalks, lighting…), interpretations and a scene summary. All numbers use the recommended prompt (v5.2c), greedy decoding and the same photos for every model. "Generic NVFP4" is a publicly distributed 4-bit version…

This model is a fine-tuned version of Qwen/Qwen3.8-27B on the on the Omni-Edu-70K dataset. The following hyperparameters were used during training: - learningrate: 5e-06 - trainbatchsize: 1 - evalbatchsize: 8 - distributedtype: multi-GPU - numdevices: 16 - gradientaccumulationsteps: 8 - totaltrainbatchsize: 128 - totalevalbatchsize: 128 - lrschedulertype: cosine - lrschedulerwarmupsteps: 0.1 - numepochs: 3.0 - Transformers 5.2.0 - Pytorch 2.10.0 - Datasets 4.0.0 - Tokenizers 0.22.2

Open Foundation Models for Learning and Teaching OmniEdu-27B-FP8 is the FP8-quantized release of OmniEdu-27B. The architecture, tokenizer, chat template and instruction-tuning data are unchanged: this checkpoint is the same model with its linear layers stored in FP8, which shrinks the weights from 50.96 GiB to 28.30 GiB and is intended for deployment on smaller GPUs and for higher-throughput serving. This checkpoint accompanies OmniEdu: Open Foundation Models for Learning and University · University of the Chinese Academy of Sciences · Zhongguancun Academy Quantized with LLM Compressor into the compressed-tensors FP8BLOCK scheme (library version 0.19.0), without calibration data: The vision…

Open Foundation Models for Learning and Teaching OmniEdu-4B-FP8 is the FP8-quantized release of OmniEdu-4B. The architecture, tokenizer, chat template and instruction-tuning data are unchanged: this checkpoint is the same model with its linear layers stored in FP8, which shrinks the weights from 8.46 GiB to 5.13 GiB and is intended for deployment on smaller GPUs and for higher-throughput serving. This checkpoint accompanies OmniEdu: Open Foundation Models for Learning and University · University of the Chinese Academy of Sciences · Zhongguancun Academy Quantized with LLM Compressor into the compressed-tensors FP8BLOCK scheme (library version 0.19.0), without calibration data: The vision…