TRACE is a trajectory-aware defense model for multi-turn jailbreaks. Instead of judging each user message in isolation, it commits to an explicit safety assessment of the whole conversation so far — in a block — and only then writes its reply in an block. The assessment is part of the generation, so the reply is conditioned on it. This is the Qwen3-8B member of the TRACE family; see Dipto084/Llama3.1-8B-TRACE for the Llama-3.1-8B counterpart. This is the model-agnostic transfer experiment of the TRACE paper (§6.3): the full recipe — SFT followed by GRPO, every component of the pipeline held fixed — applied to a model from a different architecture family, with a distinct base…
SAVRN Model Hub · Models by Task
Text Generation Models
1,582 open-weight text generation models in the SAVRN Model Hub, with Convergent Intelligence, Qwen and Francesca Padovani publishing the most.
1,582Models
732Publishers
107K to 1.7TParameter range
8Licenses
SAVRN's Take
Text generation is where most of the attention goes, and where most of the money is spent. The models on this page run from tiny research checkpoints that fit on a laptop to frontier-class systems that need a full server of accelerators. The first question we ask about any of them is not how smart it is, but what it costs to keep it answering around the clock.
That cost comes down to three things: the size of the weights, the precision you run them at, and how much context each request carries. Moving from 16-bit to 4-bit cuts the memory a model needs by roughly four times, which is often the difference between one server and several. Mixture-of-experts models, such as the newer DeepSeek, Qwen and Nemotron releases, hold a large total parameter count but activate only a small share per token, so they answer faster than their size suggests while still needing the memory to hold every expert.
Licensing decides the rest. Apache 2.0 and MIT models can go straight into a commercial product. Llama, Gemma's earlier generations and models marked other each carry terms you need to read first, and a model with no stated license is research material until the publisher says otherwise. Our advice is to shortlist by license, size to the hardware you can actually run, then test the survivors on your own prompts. The table below is sorted by what people download, not by what we recommend.
Most Downloaded
| Model | Publisher | Parameters | License | Monthly downloads | Cheapest GPUs at 16-bit |
|---|---|---|---|---|---|
| Qwen3-0.6B | Qwen | 752M | apache-2.0 | 25.2M | 1x MI300X, $1.85/hr |
| gpt2 | OpenAI community | 137M | mit | 18.8M | 1x MI300X, $1.85/hr |
| Qwen3-4B | Qwen | 4B | apache-2.0 | 11.4M | 1x MI300X, $1.85/hr |
| Qwen3-14B | Qwen | 14.8B | apache-2.0 | 9.4M | 1x MI300X, $1.85/hr |
| Qwen3-8B | Qwen | 8.2B | apache-2.0 | 8.1M | 1x MI300X, $1.85/hr |
| Qwen2.5-7B-Instruct | Qwen | 7.6B | apache-2.0 | 6.4M | 1x MI300X, $1.85/hr |
| gpt-oss-20b | OpenAI | 20.9B | apache-2.0 | 5.7M | 1x MI300X, $1.85/hr |
| opt-125m | AI at Meta | — | other | 5.6M | — |
| DeepSeek-V3.2 | DeepSeek | 685.4B | mit | 5.4M | 7x MI325X, $14.00/hr |
| Qwen2.5-0.5B-Instruct | Qwen | 494M | apache-2.0 | 5.3M | 1x MI300X, $1.85/hr |
Licenses
| License | Models | Commercial use |
|---|---|---|
| apache-2.0 | 734 | Yes |
| not stated | 345 | Not stated |
| other | 175 | Read the license |
| mit | 152 | Yes |
| cc-by-nc-4.0 | 42 | Not without separate permission |
| llama3.2 | 33 | Read the license |
| gemma | 22 | Yes, with conditions |
| llama3.1 | 21 | Yes, with conditions |
Who Publishes Them
| Publisher | Models |
|---|---|
| Convergent Intelligence | 54 |
| Qwen | 42 |
| Francesca Padovani | 37 |
| Yanggangu | 28 |
| Shreyansh singh | 25 |
| Kumar | 23 |
All 1,582 Models, Page 10 of 27
ProseLens detects who wrote the words of a document. It is IdeaLens's counterpart in the paper: the same backbone, training documents and labels, but it reads the raw document text instead of an outline, so it learns word-level provenance. It returns P(human), the probability that the document was written by a person. ProseLens is nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-BF16 fine-tuned with LoRA (rank 64) on 1M English web documents (WildOutlines). Try it in your browser: the IdeaLens & ProseLens demo scores your own text with both detectors, no installation or keys needed. From the paper: ProseLens is accurate when a document's ideas and words come from the same source (99.1%), but…
This version of Qwen3-1.7B-GPTQ-Int4 has been converted to run on the Axera NPU using w4a16 quantization. This model has been optimized with the following LoRA: For those who are interested in model conversion, you can try to export axmodel through the original repo: https://huggingface.co/Qwen/Qwen3-1.7B Convert the original Huggingface Qwen3-1.7B-GPTQ-Int4 to axmodel, and then apply the w4a16 quantization to get the final axmodel for axllm runtime. 方式二：一行命令安装（默认分支 axllm）： 方式三：下载Github Actions CI 导出的可执行程序（适合没有编译环境的用户）： https://github.com/AXERA-TECH/ax-llm/actions?query=branch%3Aaxllm 下载 最新 CI 导出的可执行程序（axllm），然后：
Run Google's Gemma 4 12B IT locally on an Apple silicon Mac with Core ML. One int4 bundle supports conversations up to 131,072 tokens, switching from its 32K mode to its 128K mode without reloading weights or copying the conversation cache. Requires: macOS 26+ · Xcode 26 / Swift 6.2 · ~11 GB disk · ~12 GB inference memory, 24 GB+ recommended. See Requirements for measured peaks. The script downloads this bundle to./models/gemma-4-12b-it-coreml-128k. The model guide also covers one-shot prompts and CLI flags. For the app, open CoreMLSamples.xcodeproj at the repository root and run the CoreMLSamples scheme on My Mac. For a standalone download, the existing Hub command is: The runtime supplies…
This model is currently still in post-training and is pre-released as a preview checkpoint with gated access only.
基于 Qwen2.5-7B 的中医领域继续预训练与监督微调模型。 - 基座模型：Qwen/Qwen2.5-7B - 架构：Qwen2 - 训练方式：中医 CPT + SFT - 权重格式：Safetensors 本模型以 Qwen/Qwen2.5-7B 为基座。使用和分发时请同时遵守基座模型及相关训练数据的适用许可与条款。
Author: Bhaskar Tallamraju and Vishal Saxena, Privify Inc. · Code and benchmark: github.com/Privify-Inc/aiact-tier · Fine-tuned from: Qwen3-4B (Alibaba Cloud / Qwen Team, Apache 2.0) · License: Apache 2.0. The base model's capabilities come from the Qwen team's work; please cite the Qwen3 Technical Report alongside this model. A 4B-parameter model, fine-tuned from Qwen3-4B with LoRA (Low-Rank Adaptation, a lightweight fine-tuning method), that classifies a described AI system under the EU AI Act (Regulation (EU) 2024/1689). It returns a risk tier and the provisions that apply, as JSON. Provided as is, with no warranty and no liability (Apache 2.0, sections 7 and 8). Not legal advice, not a…
Fixed memory for long-running generation. This EdgeRock-4B deployment package retains recurrent state, local attention, and Rock Memory Anchors (RMA), keeping persistent history at a fixed capacity as context grows. 固定记忆，长程生成。 本版本保留 EdgeRock-4B 的递推状态、局部注意力和记忆岩锚（RMA），历史内存不随上下文增长。 Rock Matrix Whitepaper / 岩芯矩阵技术白皮书： whitepaper connects EdgeRock's fixed memory, DiffusionRock's parallel generation, and SiliconRock's model-native silicon into the route toward Ultra-Fast Token. 白皮书讲清固定记忆、并行生成、模型固化三条路线，以及它们如何走向更快、更经济的生成。 Public development preview for Google AI Edge Gallery device validation. The container includes the text decoder, tokenizer, ChatML prompt template and fixed-state executor…
Gleipnir 4B is a research monitor for AI control: it scores visible trajectories of tool-using AI agents for sabotage, data exfiltration, and other covert unauthorized actions taken while appearing to do an assigned task. It is a LoRA adapter for Qwen/Qwen3.5-4B, not a standalone model. Version 0.2 is trained on agent tool-use trajectories only, with the same data and recipe as Gleipnir 27B. It is the exact weight average of three training seeds (see Training). This release is an early research artifact intended for reproducibility and follow-up work. There is not yet a Gleipnir paper or blog post. It should not be treated as a production safety system or as evidence that an interaction is…
This repository contains the trained PEFT LoRA adapter for DuoNeural/LFM2.5-8B-A1B-Hermes-Agentic-Coder-Abliterated. Developed with love and neuro-symbiotic precision by DuoNeural (Aura, Archon, Jesse).
This is the LoRA adapter form (40 MB) of lododo-Distill-v1.0. The model itself: a distilled fine-tune of the Qwen3.6-35B-A3B base (MoE, 3B active), trained on data generated by Qwen3.8-27B as the teacher — bringing the base to (and beyond) the teacher's capability while keeping the MoE speed advantage. moefast is a small launcher that configures llama.cpp for your machine. It reads the model's GGUF metadata, measures your free VRAM, computes the expert-cache size, context and batch parameters, writes them into an editable config file next to the model, and starts llama-server. The expert cache itself — keeping expert weights in host RAM and caching the hot ones in VRAM — is llama.cpp's own…
CHILLGuard: Towards Fine-Grained Chinese LLM Safety Guardrail with Scalable Data Construction and Model-aware Preference Alignment (Findings of EMNLP 2026) CHILLGuard-8B is a fine-grained Chinese LLM content-safety guardrail classifier fine-tuned from Qwen3-8B. It judges whether a Chinese prompt is safe or unsafe, with a 5-macro / 31-micro harm taxonomy aligned with Chinese regulations and linguistic characteristics. - Achieves state-of-the-art performance on Chinese content-safety benchmarks (see the paper for detailed evaluation results). The model is trained with the generator–classifier collaborative training framework via Model-aware Direct Preference Optimization (MDPO) described in…
TRACE is a trajectory-aware defense model for multi-turn jailbreaks. Instead of judging each user message in isolation, it commits to an explicit safety assessment of the whole conversation so far — in a block — and only then writes its reply in an block. The assessment is part of the generation, so the reply is conditioned on it. This checkpoint is the paper's TRACE-GRPO model: a LoRA policy trained with group-decoupled GRPO on top of the stage-1 SFT adapter, merged into the base weights for release. Across seven multi-turn attack benchmarks it averages 14.5% ASR, against 31.4% for the strongest baseline and 74.9% for the undefended target, while keeping 93.3% average compliance on…
Modelos afinados (fine-tuned) para operar como el cerebro autónomo de OzyAssist en Windows. - Eliminación Total de Alucinaciones: Cero bucles de saltos de línea (\n\n\n) ni copias textuales del System Prompt.
ProKope-7B is the #1 ranked open-weight reasoning model in the world in the 7B parameter class on the official Hugging Face Open LLM Leaderboard v2 (Evaluation Snapshot: October 2026). Developed by ProKope AI, ProKope-7B was engineered to solve one of the fundamental bottlenecks of sub-10B language models: achieving elite multi-step deductive reasoning, graduate-level scientific problem-solving, and robust instruction compliance without suffering from catastrophic reasoning collapse or narrow benchmark memorization. Evaluating across EleutherAI's rigorous lm-evaluation-harness Leaderboard v2 gauntlet, ProKope-7B achieves an official 44.75% Composite Score, setting the highest benchmarked…
A 15,070,848-parameter Japanese decoder-only language model for local contextual candidate reranking and short within-sentence continuation. This release contains the selected V4.2 epoch 20 / step 1,081,942 FP32 inference bundle, its INT8 block32, FP32-compute Core ML stateless package, and a matching chunked KV-cache Core ML package. Status: deployed default. V4.2 is the default reranking and next-word model of the Vime input method; V2.1 ships alongside it as the fallback. V4.2 is not a demonstrated accuracy improvement over V2.1 (see Evaluation). This is a custom PyTorch model, not a Transformers AutoModel integration or an instruction/chat model. SentencePiece, kana retrieval, candidate…
Bielik-11B-v3.0-Instruct post-trained for the Polish history matura (extended level) during the Warsaw Model Trainers alpha=64) on top of speakleash/Bielik-11B-v3.0-Instruct. closedem = exact match on closed items. Graded scores (deepseek-flash-high@think via benchmark/bench.py) are reported separately.
CHILLGuard: Towards Fine-Grained Chinese LLM Safety Guardrail with Scalable Data Construction and Model-aware Preference Alignment (Findings of EMNLP 2026) CHILLGuard-1.7B is a fine-grained Chinese LLM content-safety guardrail classifier fine-tuned from Qwen3-1.7B. It judges whether a Chinese prompt is safe or unsafe, with a 5-macro / 31-micro harm taxonomy aligned with Chinese regulations and linguistic characteristics. - Achieves state-of-the-art performance on Chinese content-safety benchmarks (see the paper for detailed evaluation results). The model is trained with the generator–classifier collaborative training framework via Model-aware Direct Preference Optimization (MDPO) described…
CHILLGuard: Towards Fine-Grained Chinese LLM Safety Guardrail with Scalable Data Construction and Model-aware Preference Alignment (Findings of EMNLP 2026) CHILLGuard-4B is a fine-grained Chinese LLM content-safety guardrail classifier fine-tuned from Qwen3-4B. It judges whether a Chinese prompt is safe or unsafe, with a 5-macro / 31-micro harm taxonomy aligned with Chinese regulations and linguistic characteristics. - Achieves state-of-the-art performance on Chinese content-safety benchmarks (see the paper for detailed evaluation results). The model is trained with the generator–classifier collaborative training framework via Model-aware Direct Preference Optimization (MDPO) described in…
A 75M parameter decoder-only language model built entirely from scratch in PyTorch — no HuggingFace model classes, no nanoGPT wrapping. Every component (tokenizer, architecture, data pipeline, training loop, SFT) was written from scratch with Claude (Anthropic's AI assistant). Pretraining - ~17.8B tokens of English web text - AdamW (β₁=0.9, β₂=0.95), weight decay 0.1 SFT Fine-tuning - 100K examples from OpenHermes-2.5 - ChatML format with loss masking on user/system tokens Evaluated with log-likelihood scoring (no few-shot): Comparable to GPT-2 (117M) at 0.64× the parameter count. This is a small research model, built to learn how language models work from the ground up. It is not suitable…
A 12,537,920-parameter Japanese decoder-only language model for local contextual candidate reranking and short within-sentence continuation. This release contains the selected V2.1 extend5 best / step40000 FP32 inference bundle and its INT8 block32, FP32-compute Core ML variant. V2 experiments completed on 2026-10-08; the current quantization loss and observed input latency are accepted for this deployment. This is a custom PyTorch model, not a Transformers AutoModel integration or an instruction/chat model. SentencePiece, kana retrieval, candidate scoring and search run outside the Core ML graph. The original architecture, loader and scoring code are included under source/; no…
RoPE-free long-range attention, trained at 4k with zero-shot length extrapolation. FS-B searches full-prefix and norm-filtered forests of chunk mean keys, reranks their candidates, and reads four chunks. Its attention is superlinear and subquadratic. This 743M research base model was trained on FineWeb-Edu at a 4,096-token context. The short retrieval example asks for the hidden number 5437923. Start with CUDA-enabled PyTorch; Usage gives the tested installation and speed-replication command. FreeSolo learns simple chunk mean keys that support search, norm pruning, and hierarchical summarization. RoPE stays in the local window; long-range attention reads selected original tokens without…
RoPE-free long-range attention, trained at 4k with zero-shot length extrapolation. FS-L retains a bounded pool of chunks by key norm and recency, then reads four selected chunks. Its attention scales linearly with sequence length. This 743M research base model was trained on FineWeb-Edu at a 4,096-token context. The short retrieval example asks for the hidden number 5437923. Start with CUDA-enabled PyTorch; Usage gives the tested installation and speed-replication command. FreeSolo learns simple chunk mean keys that support search, norm pruning, and hierarchical summarization. RoPE stays in the local window; long-range attention reads selected original tokens without rotating their queries…
LoRA adapter for google/gemma-4-12B-it-qat-q40-unquantized trained to answer the Polish history matura (poziom rozszerzony, CKE). Built at the Warsaw Model Trainers Hackathon (team repo: WarsawModelTrainersHackathon, src/posttrain/). Supervised fine-tuning, bf16 LoRA on the QAT weights (deployed on the official Q40 GGUF). One mixed pass over every source: short-key items (real/rephrased/synthetic exam items, textbook questions) trained with thinking OFF (question -> CKE answer), gemma-4-31b@think teacher traces (judge-filtered, exam items + essays) trained with thinking ON; exam-like rows oversampled x2-x3. Loss on the answer only. LoRA r=32, alpha=64, all language-model linear layers, lr…
LoRA adapter for google/gemma-4-12B-it-qat-q40-unquantized trained to answer the Polish history matura (poziom rozszerzony, CKE). Built at the Warsaw Model Trainers Hackathon (team repo: WarsawModelTrainersHackathon, src/posttrain/). Vision SFT (QLoRA, 4-bit base): image-based exam items (real CKE papers with their source images, synthetic Wikimedia image tasks, synthetic full papers) with the PNGs fed to the model, plus 30% of the teacher traces; vision tower frozen, language LoRA only. LoRA r=32, alpha=64, lr 1e-4, 1 epoch, batch 1 x gradaccum 16, maxlen 6144, Unsloth vision collator Consolidated training set data/trainingdatasets/sfttrain.jsonl (buildsft.py): real CKE papers 2005–2026…
Japanese description is available below. This model is a fully merged model created by taking llm-jp/llm-jp-4-8b-thinking as the base model. It is fine-tuned using QA data containing both Imabari dialect and standard Japanese reasoning text and answers. A custom chat template provides the analysisimabari and finalimabari channels alongside the standard analysis and final channels. In v3, there was an issue where sentences would repeat indefinitely with reasoningeffort="high". Although this issue has been partially mitigated, repetitive output now occasionally occurs even with reasoningeffort="medium", so the issue has not yet been fully resolved. This repository contains the merged full…
An 11 GB GGUF of Sarvam-30B that runs on a CPU with about 12 GB of free RAM. It scores 81.7% on MMLU, against 85.1 published for the full model. Sarvam-30B is a mixture-of-experts model: 32 billion weights, about 2.4 billion of them active per token, which is why a CPU can run it. Use the IQ2M file. The IQ1M file is here for comparison: it saves 3 GB but loses much more quality. For the larger Sarvam-105B, see Aaravomega/sarvam-105b-GGUF-IQ2M-IQ1M (2-bit and 1-bit) and Aaravomega/sarvam-105b-GGUF-Q80 (8-bit). You need a recent llama.cpp build with Sarvam support (architecture bailingmoe2; tested with release b11731) and about 12 GB of free RAM. Then open http://127.0.0.1:8080, or call the…
A frozen 7,386,624-parameter, decoder-only Japanese language model for local contextual candidate reranking and short within-sentence continuation. This repository provides the FP32 inference bundle and an iOS18 INT8-weight Core ML variant. SentencePiece tokenization, kana candidate retrieval, scoring and search run outside the Core ML graph. This is a custom PyTorch model, not a Transformers AutoModel integration or an instruction/chat model. The source model, loader, scoring and conversion code are included under source/; no trustremotecode or training checkpoint is needed. The original manifests remain byte-for-byte intact. Their old absolute paths describe provenance, not paths required…
People's InvoiceLM-14B is a domain-specific fine-tuned model built on Qwen3-14B base model, optimised for template-free invoice understanding, structured data extraction, and real-world finance document workflows. It is designed for the messy reality of enterprise invoices: changing supplier formats, inconsistent field layouts, scanned documents, and operational data that must be reviewed before it reaches downstream systems. People's InvoiceLM powers invoice intelligence capabilities within E-Flow, Effectz.AI's document execution platform. People's LM is Effectz.AI's family of compact, specialised language models built for real-world business workflows in emerging markets — built for the…
This repository contains experimental Group Relative Policy Optimization (GRPO) LoRA weights trained to evaluate test-time compute and reinforcement learning on mathematical, physical, and scientific reasoning. Test this model instantly on Google Colab with an embedded Gradio chat interface: Released under the Creative Commons Attribution-NonCommercial 4.0 International (CC BY-NC 4.0) license. Academic research, university coursework, student study, and personal experimentation. Any commercial deployment, paid SaaS API, commercial tutoring platform, or enterprise wrapper requires an explicit commercial license. This architecture is formally documented and validated in the Vigyan AI…
This repository contains an experimental domain-specialized LoRA adapter trained as part of the Vigyan AI Sovereign Mixture of Experts (MoE) Upcycling Initiative. The goal of this experimental line was to train isolated, high-rank domain experts on specialized STEM corpora and examine whether post-hoc upcycling via uncalibrated linear routers (mergekit-moe) could synthesize dense reasoning models into edge-deployable sparse MoE networks. During post-training MoE upcycling experiments, this adapter was used in a 4-expert sparse mixture configuration. The experimental run revealed pivotal architectural insights: 1. Attention-FFN Desynchronization: Because this LoRA was trained across both…
This repository contains an experimental domain-specialized LoRA adapter trained as part of the Vigyan AI Sovereign Mixture of Experts (MoE) Upcycling Initiative. The goal of this experimental line was to train isolated, high-rank domain experts on specialized STEM corpora and examine whether post-hoc upcycling via uncalibrated linear routers (mergekit-moe) could synthesize dense reasoning models into edge-deployable sparse MoE networks. During post-training MoE upcycling experiments, this adapter was used in a 4-expert sparse mixture configuration. The experimental run revealed pivotal architectural insights: 1. Attention-FFN Desynchronization: Because this LoRA was trained across both…
This repository contains an experimental domain-specialized LoRA adapter trained as part of the Vigyan AI Sovereign Mixture of Experts (MoE) Upcycling Initiative. The goal of this experimental line was to train isolated, high-rank domain experts on specialized STEM corpora and examine whether post-hoc upcycling via uncalibrated linear routers (mergekit-moe) could synthesize dense reasoning models into edge-deployable sparse MoE networks. During post-training MoE upcycling experiments, this adapter was used in a 4-expert sparse mixture configuration. The experimental run revealed pivotal architectural insights: 1. Attention-FFN Desynchronization: Because this LoRA was trained across both…
This repository contains an experimental domain-specialized LoRA adapter trained as part of the Vigyan AI Sovereign Mixture of Experts (MoE) Upcycling Initiative. The goal of this experimental line was to train isolated, high-rank domain experts on specialized STEM corpora and examine whether post-hoc upcycling via uncalibrated linear routers (mergekit-moe) could synthesize dense reasoning models into edge-deployable sparse MoE networks. During post-training MoE upcycling experiments, this adapter was used in a 4-expert sparse mixture configuration. The experimental run revealed pivotal architectural insights: 1. Attention-FFN Desynchronization: Because this LoRA was trained across both…
Quantized builds of Qwen/Qwen3.8-27B in three sizes. On held-out text, each one is closer to the BF16 model than the unsloth dynamic quant of the same size. Each GGUF is no larger than its unsloth counterpart. The Small bundle needs baseRT 0.4.0 or later. KL divergence of each quant's next-token distribution from the BF16 model, ours beside unsloth's quant of the same size, measured in llama.cpp. Lower is closer to BF16. The percentage over each pair is our change against unsloth. On text that was not used to build these files, all three are 17–34% closer to BF16 than unsloth's quant of the same size. Each cell: unsloth KL / BaseQuant KL (change). Bold marks the lower KL. The prompts come…
本仓库提供由 Qwen/Qwen3-0.6B 转换得到的 RKNN LLM 模型。 - Model ID：RKNNAI/RK182X-LLM-Qwen3-0.6B - 模型显示名称：RK182X-LLM-Qwen3-0.6B - 源模型：Qwen/Qwen3-0.6B - 模型类型：LLM ModelScope 完整下载： Hugging Face 完整下载： ModelScope 指定配置下载： Hugging Face 指定配置下载： - 使用配套 RKNN Runtime 和驱动；运行前用 rknn-smi -v 检查设备端版本。 源模型许可证见根目录 LICENSE；同时遵守 RKNN Toolkit 和 RKNN Runtime 许可条款。
本仓库提供由 Qwen/Qwen3.5-0.8B 转换得到的 RKNN LLM 模型。 - Model ID：RKNNAI/RK182X-LLM-Qwen3.5-0.8B - 模型显示名称：RK182X-LLM-Qwen3.5-0.8B - 源模型：Qwen/Qwen3.5-0.8B - 模型类型：LLM ModelScope 完整下载： Hugging Face 完整下载： ModelScope 指定配置下载： Hugging Face 指定配置下载： - 使用配套 RKNN Runtime 和驱动；运行前用 rknn-smi -v 检查设备端版本。 源模型许可证见根目录 LICENSE；同时遵守 RKNN Toolkit 和 RKNN Runtime 许可条款。
本仓库提供由 Qwen/Qwen3.5-2B 转换得到的 RKNN LLM 模型。 - Model ID：RKNNAI/RK182X-LLM-Qwen3.5-2B - 模型显示名称：RK182X-LLM-Qwen3.5-2B - 源模型：Qwen/Qwen3.5-2B - 模型类型：LLM ModelScope 完整下载： Hugging Face 完整下载： ModelScope 指定配置下载： Hugging Face 指定配置下载： - 使用配套 RKNN Runtime 和驱动；运行前用 rknn-smi -v 检查设备端版本。 源模型许可证见根目录 LICENSE；同时遵守 RKNN Toolkit 和 RKNN Runtime 许可条款。
Vigyan-3B 4× MoE is an experimental Sparse Mixture of Experts model upcycled from 4 specialized domain LoRA adapters (Science, Technology, Engineering, Mathematics). It was assembled to evaluate whether post-hoc expert stitching on compact language models ($\le 3\text{B}$) can deliver multi-domain specialization at single-expert inference latency. - shreyansh12183/vigyan-3b-adapter-science - shreyansh12183/vigyan-3b-adapter-technology - shreyansh12183/vigyan-3b-adapter-engineering - shreyansh12183/vigyan-3b-adapter-mathematics 1. Delimiter Blindness: Without general conversational anchor replay during adapter training, the model degrades into token echo loops when receiving general chat…
Community GGUF quantizations of whitecircle/GLM-4.7-Flash-Coder. 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. GLM-4.7-Flash-Coder is an agentic coding fine-tune of zai-org/GLM-4.7-Flash. The upstream model card describes a 30B Mixture-of-Experts model trained on software-engineering trajectories, with tool use, reasoning, and concurrent tool-call behavior. The upstream training configuration uses a 75k-token sequence length; its model configuration declares a 202,752-token maximum position limit. This release does not claim long-context behavior beyond the runtime…
Qwen3.8-Flash-Next (125B total, 6B active) in the oom-inference weight format, for serving on one 24 GB GPU with 64 GB of RAM. The engine keeps routed experts across VRAM, pinned RAM and NVMe, so the model does not need to fit in memory. This is with its bytes laid out one record per expert. No weight is changed: the NVFP4 routed experts, BF16 dense weights and FP8 n-gram tables are exactly as released. It is only for oominf; it is not a Transformers, Safetensors or GGUF checkpoint. How it works: How to serve 134 GB of model weights with 24 GB VRAM / 64 GB RAM. Measured through oominf serve's OpenAI-compatible API with oomeval, on the files in this repository with serve defaults (fp32…
本仓库提供由 google/functiongemma-270m-it 转换得到的 RKNN LLM 模型。 - Model ID：RKNNAI/RK182X-LLM-functiongemma-270m-it - 模型显示名称：RK182X-LLM-functiongemma-270m-it - 源模型：google/functiongemma-270m-it - 模型类型：LLM ModelScope 完整下载： Hugging Face 完整下载： ModelScope 指定配置下载： Hugging Face 指定配置下载： - 使用配套 RKNN Runtime 和驱动；运行前用 rknn-smi -v 检查设备端版本。 源模型许可证见根目录 LICENSE；同时遵守 RKNN Toolkit 和 RKNN Runtime 许可条款。 本发布物为源模型的 RKNN 转换/量化衍生物，保留原始许可与归属声明，见 Notice。 源模型使用 Gemma Terms of Use，并非 Apache 2.0。使用与再分发同时受 使用限制 约束。
本仓库提供由 LiquidAI/LFM2.5-1.2B-Instruct 转换得到的 RKNN LLM 模型。 - Model ID：RKNNAI/RK182X-LLM-LFM2.5-1.2B-Instruct - 模型显示名称：RK182X-LLM-LFM2.5-1.2B-Instruct - 源模型：LiquidAI/LFM2.5-1.2B-Instruct - 模型类型：LLM ModelScope 完整下载： Hugging Face 完整下载： ModelScope 指定配置下载： Hugging Face 指定配置下载： - 使用配套 RKNN Runtime 和驱动；运行前用 rknn-smi -v 检查设备端版本。 - 需使用支持 shortconv pads=2 的配套推理实现。 源模型许可证见根目录 LICENSE；同时遵守 RKNN Toolkit 和 RKNN Runtime 许可条款。 本发布物为源模型的 RKNN 转换/量化衍生物，保留原始许可与归属声明，见 Notice。 源模型使用 LFM Open License v1.0，并非 Apache 2.0。
本仓库提供由 Qwen/Qwen2.5-1.5B-Instruct 转换得到的 RKNN LLM 模型。 - Model ID：RKNNAI/RK182X-LLM-Qwen2.5-1.5B-Instruct - 模型显示名称：RK182X-LLM-Qwen2.5-1.5B-Instruct - 源模型：Qwen/Qwen2.5-1.5B-Instruct - 模型类型：LLM ModelScope 完整下载： Hugging Face 完整下载： ModelScope 指定配置下载： Hugging Face 指定配置下载： - 使用配套 RKNN Runtime 和驱动；运行前用 rknn-smi -v 检查设备端版本。 源模型许可证见根目录 LICENSE；同时遵守 RKNN Toolkit 和 RKNN Runtime 许可条款。 本发布物为源模型的 RKNN 转换/量化衍生物，保留原始许可与归属声明，见 Notice。
本仓库提供由 Qwen/Qwen2.5-3B-Instruct 转换得到的 RKNN LLM 模型。 - Model ID：RKNNAI/RK182X-LLM-Qwen2.5-3B-Instruct - 模型显示名称：RK182X-LLM-Qwen2.5-3B-Instruct - 源模型：Qwen/Qwen2.5-3B-Instruct - 模型类型：LLM ModelScope 完整下载： Hugging Face 完整下载： ModelScope 指定配置下载： Hugging Face 指定配置下载： - 使用配套 RKNN Runtime 和驱动；运行前用 rknn-smi -v 检查设备端版本。 源模型许可证见根目录 LICENSE；同时遵守 RKNN Toolkit 和 RKNN Runtime 许可条款。 本发布物为源模型的 RKNN 转换/量化衍生物，保留原始许可与归属声明，见 Notice。 源模型使用 Qwen RESEARCH LICENSE AGREEMENT，并非 Apache 2.0。 Built with Qwen.
本仓库提供由 Qwen/Qwen3-4B-Instruct-2507 转换得到的 RKNN LLM 模型。 - Model ID：RKNNAI/RK182X-LLM-Qwen3-4B-Instruct-2507 - 模型显示名称：RK182X-LLM-Qwen3-4B-Instruct-2507 - 源模型：Qwen/Qwen3-4B-Instruct-2507 - 模型类型：LLM ModelScope 完整下载： Hugging Face 完整下载： ModelScope 指定配置下载： Hugging Face 指定配置下载： - 使用配套 RKNN Runtime 和驱动；运行前用 rknn-smi -v 检查设备端版本。 源模型许可证见根目录 LICENSE；同时遵守 RKNN Toolkit 和 RKNN Runtime 许可条款。
A little over a month has passed since Spark-X2.5. In the rapidly evolving world of AI agents, that is enough time to rethink, refine, and move the frontier forward. Building on the Spark-X2.5-4B, we introduce Spark-X2.6-4B, the latest generation of our compact, general-purpose language models for on-device deployment. Spark-X2.6-4B retains exactly the same architecture as Spark-X2.5-4B, achieving its capability gains through improved post-training. It delivers substantial improvements in coding, agent, long-context, multilingual, reasoning, etc. These advances enable more capable and practical on-device AI for developers and users tackling real-world tasks. - Stronger Coding, Agentic, and…
GGUF quants of Step-5-Preview — a 604B-class sparse MoE (measured 598.04 B parameters across 1,548 tensors), 352 routed experts with top-8 routing, ~16B active, 95 layers (92 text + 3 MTP), 128,896 vocab, 1,048,576 context. This repo is private and is not a general release. The quant packs carry the text trunk — 598.04 B of the checkpoint's 602.97 B parameters. The remaining 4.9325 B ship alongside them as GGUF companion files; see Companion files. They sit outside the packs because llama.cpp's step35 implements no runtime for them, not because they were dropped. One folder per quant type, with the quant type in every file name — the convention unsloth uses, so the Hub and every downstream…
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).
This repository collects Yihan's 24 research runs: six injection doses (10, 25, 50, 100, 250, 500) and four source allocations (M, C, G, MCG). The exact trigger is dAtA[iNdEx]. Models use the Pythia-160M architecture and train from random initialization, seed 42, for 16,000,090,112 tokens on the frozen Dolma Math/Code/General mixture. The campaign is in progress. A run is released only when its checkpoint audit and all scheduled evaluations pass. releases/.json records uploaded file checksums and the verified Hub revision. Missing runs are pending, not zero-valued results. The repository is public, matching the group's reference release. Scheduled checkpoints may appear before a run is…
Weights of 3D-PLOT-LLM: Part-Level Object Tokens for 3D Large Language Models (NeurIPS 2026). 3D-PLOT-LLM extends PointLLM (Vicuna-v1.5-7B with a frozen Point-BERT encoder on 8192-point Objaverse clouds). The encoder's patch tokens are partitioned into K=16 regions; each region is prefixed with a learnable marker and a reserved vocabulary token, refined by a lightweight Marker-Space Refinement module. The model can answer which regions a part description refers to, describe a given set of regions, and caption whole objects, with fewer than one million added parameters. The checkpoint is in PointLLM format (PointLLMLlamaForCausalLM, bf16) and loads with the pinned transformers commit used by…
This repository contains an instruction-tuned causal language model for text generation and chat-style prompts. Use a recent version of transformers that supports this model architecture. The following example loads the model from this repository and generates a response.
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).
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).
Three LoRA adapters trained jointly in Stage 1 of the CoTMTKD pipeline. Each adapter is one "expert" on top of Qwen/Qwen2.5-7B-Instruct; the experts are trained with an SFT loss plus a DPP repulsion term that pushes them apart. They are meant to be used as teachers for the Stage 2 distillation step of that pipeline. The pipeline's own final/adapterstates.pt bundle and the resumable checkpoint.pt referenced in manifest.json are not uploaded. Absolute local paths in config.yaml and manifest.json were rewritten as relative paths, so the configfilesha256 in the manifest no longer matches config.yaml. - 3 epochs, 94 optimizer steps, global batch size 32, learning rate 5e-5, AdamW, cosine…
"Abliterated Dolphin" is a result of my 3AM brain reading about technique called abliteration and then thinking what would happen if I tried to abliterate a model that is already relatively free, such as Dolphin. Heavily inspired by mlabonne's article on abliteration on how to redirect refusals and effectively remove, or ablate, censorship from a language model. There is really no deeper meaning to any of this than pure curiosity. This is basically a dumbed down version of the original Dolphin model I used as a base, as I have not done any DPOs to heal the damage caused by abliteration. Don't try to do anything meaningful with this model. Use the original Dolphin3.0-R1-Mistral-24B instead.…
ACE-3-26B-A4B-Preview-260910 is a preview release of APMIC's ACE-3 model family, built for Traditional Chinese (Taiwan) enterprise scenarios and agentic workflows. The model is based on google/gemma-4-26B-A4B-it, a Mixture-of-Experts model with 26B total parameters and about 4B active parameters per token (the "A4B" in the name). APMIC has further optimized it to strengthen: - Traditional Chinese output in Taiwan usage (terminology, phrasing and orthography) The bundled generationconfig.json uses temperature=1.0, topp=0.95, topk=64. These follow the base model's recommended settings. The model can be served with any inference engine that supports Gemma 4 (for example vLLM). Because the…
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).
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).
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).
Questions
Which Text generation models are most downloaded?
By monthly downloads reported by the Hugging Face Hub: Qwen3-8B-TRACE (12); ProseLens (11); Qwen3-1.7B-GPTQ-Int4 (11).