51 models · ranked by HuggingFace downloads
by Qwen
Qwen3-0.6B is the 0.6-billion-parameter instruction-tuned model from Alibaba Cloud's Qwen3 series, fine-tuned from the Qwen3-0.6B-Base for conversational and task-following use. It targets deployment in environments where even a 1B model is too large — edge hardware, mobile devices, or ultra-low-latency services. Apache 2.0 licensed.
by openai-community
OpenAI's original GPT-2 at 124M parameters, an autoregressive language model trained on WebText (over 8 million web documents filtered from Reddit outlinks). It generates English text continuation given a prompt using next-token prediction, trained without any instruction tuning or RLHF. MIT licensed and runnable on commodity CPU hardware.
by Qwen
Qwen3-8B is the 8-billion-parameter instruction-tuned model from Alibaba Cloud's Qwen3 family, positioned at the competitive midpoint between 4B and 14B+ tiers. It targets deployment on single consumer or workstation GPUs while providing strong reasoning and multilingual capabilities. Apache 2.0 licensed with text-generation-inference compatibility.
by Qwen
Qwen2.5-7B-Instruct is Alibaba Cloud's 7-billion-parameter instruction-tuned language model from the Qwen2.5 series, supporting English and a range of other languages. It targets applications requiring more reasoning and knowledge than sub-3B models, while remaining deployable on a single consumer GPU. Apache 2.0 licensed with text-generation-inference compatibility.
by nvidia
Qwen3.6-35B-A3B-NVFP4 is an NVIDIA-optimized FP4 quantization of Qwen3.6-35B-A3B, produced with the ModelOpt toolkit for deployment on NVIDIA H100/H200 GPUs. FP4 weights reduce GPU memory footprint roughly 2x compared to BF16 while maintaining most of the original accuracy for conversational tasks. It is intended for inference on NVIDIA TensorRT-LLM or vLLM backends, not for further fine-tuning.
by facebook
OPT-125M is the smallest model in Meta's Open Pretrained Transformer series, a 125-million-parameter decoder-only LLM trained on a dataset comparable to GPT-3's training mix. Released as part of Meta's effort to make large language model weights accessible for research. At 125M parameters it is primarily used for prototyping, educational purposes, and compute-constrained environments.
by Qwen
Qwen2.5-3B-Instruct is a 3-billion-parameter instruction-tuned language model from Alibaba Cloud's Qwen2.5 series, positioned between the 1.5B and 7B tiers. It targets lightweight server deployments and on-device inference scenarios where 7B is too large. The license is 'other' — requires reviewing the specific Qwen 2.5 license terms before commercial deployment.
by Qwen
Qwen2.5-1.5B-Instruct is a 1.5-billion-parameter instruction-tuned model from Alibaba Cloud's Qwen2.5 series, targeting edge and embedded deployment scenarios where even a 3B model is too large. Apache 2.0 licensed, it focuses on basic instruction following and short-context tasks at minimal compute cost.
by openai
GPT-OSS-20B is a 20-billion-parameter open-source language model released by OpenAI under Apache 2.0 — notable as OpenAI's first substantial open-weight release after years of closed-weights policy. Based on the gpt_oss architecture, it targets high-quality text generation at a scale deployable on research and enterprise GPU infrastructure. FP8 and MXfloat4 quantized variants reduce memory requirements.
by meta-llama
Llama 3.2-1B-Instruct is Meta's 1-billion-parameter instruction-tuned model from the Llama 3.2 family, the smallest Llama release targeting ultra-low-resource inference scenarios. It is designed for edge deployment on devices that cannot accommodate even 3B models. The Llama 3.2 license restricts use by products/services with over 700M monthly users.
by meta-llama
Llama 3.1-8B-Instruct is Meta's 8-billion-parameter instruction-tuned model, supporting 8 languages including English, German, French, Spanish, Italian, Portuguese, Hindi, and Thai. Released under the Llama 3.1 license (permissive with restrictions for products over 700M users), it was a leading open-weight model at its scale at release. Context window extends to 128K tokens.
by openai
OpenAI's 120B parameter open-weight language model released under Apache 2.0 in 2025. Supports MXFP4 and 8-bit quantization for multi-GPU deployment via vLLM. Competitive on reasoning and instruction-following benchmarks within the open-weight tier.
by deepseek-ai
DeepSeek-V4-Flash-0731 is a fast-inference variant of DeepSeek V4, designed to reduce time-to-first-token at the cost of some capacity. It uses the deepseek_v4 architecture with FP8 quantization (8-bit) and is MIT-licensed with official eval-results and Azure deployment support. The 2661 likes suggest it is one of the more popular recent DeepSeek releases.
by Qwen
Qwen3-4B-Instruct-2507 is a 4-billion-parameter instruction-tuned model from Alibaba Cloud's Qwen3 series, updated in July 2025. It targets the mid-range deployment tier between ultra-compact sub-2B models and the 7-8B tier requiring heavier hardware. Apache 2.0 licensed with text-generation-inference compatibility.
by Qwen
Qwen 2.5 14B Instruct is Alibaba's mid-tier instruction model with strong multilingual, coding, and math capabilities. It fills the gap between 7B-class models and the more expensive 32B/72B variants for production deployments.
by distilbert
DistilGPT2 is a knowledge-distilled version of GPT-2 small, with 82M parameters (vs GPT-2's 117M) and approximately 2x faster inference. It retains around 97% of GPT-2 small's language modeling performance while being lighter to serve.
by TinyLlama
TinyLlama 1.1B Chat is a compact instruction-tuned language model trained on 3 trillion tokens with the Llama 2 architecture. It targets deployment on devices with limited RAM while retaining basic instruction-following capability.
by deepseek-ai
DeepSeek-V3.2 is a Mixture-of-Experts (MoE) large language model from DeepSeek AI, fine-tuned from DeepSeek-V3.2-Exp-Base. It activates a subset of expert parameters per token rather than the full model, enabling high effective parameter counts at lower per-token compute cost. MIT licensed, making it freely deployable commercially despite its scale.
by meta-llama
Llama 3.2 3B Instruct is Meta's compact instruction-tuned model designed for on-device and edge inference, with strong performance for its size on reasoning and instruction following benchmarks.
by datajuicer
LLaMA-1B fine-tuned on 150B tokens of RedPajama data filtered and refined by Data-Juicer, a data-cleaning toolkit from Alibaba DAMO. The training corpus was pruned using quality heuristics across Wikipedia, arXiv, Books, and Common Crawl slices. At 1B parameters it trades capability for low inference cost.
by zai-org
GLM-5.2 is THUDM's latest iteration of the General Language Model series, featuring a sparse MoE architecture (glm_moe_dsa) with strong Chinese and English bilingual capabilities. With 4,417 likes it is one of the most widely recognized Chinese-origin open models on HuggingFace.
by mistralai
Mistral 7B Instruct v0.2 improved on v0.1 with a 32K sliding window context and better instruction following. It was the strongest 7B open-weight instruction model available when released and remains competitive for text tasks after later versions raised the bar.
by deepseek-ai
DeepSeek-R1 is a 671B parameter mixture-of-experts reasoning model from DeepSeek AI, trained with reinforcement learning to produce explicit chain-of-thought reasoning before answering. It achieves GPT-4-class performance on math, coding, and logical inference benchmarks and is released under an MIT license. Active parameters per forward pass are a subset of the 671B total, reducing compute per generated token.
by deepseek-ai
DeepSeek-R1-0528-Qwen3-8B is an 8B-parameter reasoning-focused language model built on the Qwen3 architecture, released under the MIT license. It is a distilled variant of the DeepSeek-R1 series, designed to bring chain-of-thought reasoning capabilities to a smaller, more deployable footprint. The model supports text-generation-inference and HuggingFace endpoints out of the box.
by sakamakismile
Text-only NVFP4-quantized Qwen3.6-27B with multi-token prediction (MTP) for speculative decoding, optimized for Blackwell and Hopper GPUs via NVIDIA ModelOpt. Stripping vision components reduces memory footprint and inference latency when only text output is needed. Supports 13 languages including Chinese, Japanese, and Korean.
by EleutherAI
gpt-neox-20b is EleutherAI's 20B autoregressive language model, trained on the Pile dataset and released in 2022 as the largest fully open-weights English LLM at the time. It uses the GPT-NeoX architecture with rotary position embeddings and trained in bf16 on TPUs. While now superseded by much larger models, it remains historically significant and is a baseline for open LLM research.
by openbmb
MiniCPM5-1B is OpenBMB's 1-billion parameter model from the MiniCPM5 series, targeting on-device and edge deployment. Supports long-context inference, tool calling, and Chinese/English bilingual tasks at a fraction of the compute of 7B-class models, with multiple supporting arXiv papers.
by Qwen
Qwen3-Coder 30B is a code-specialized Mixture-of-Experts model with 30B total and 3B active parameters, instruction-tuned for programming tasks. It targets agentic coding workflows including multi-file editing, tool use, and repository-level understanding.
by google
Gemma 2 2B Instruct is Google's smallest instruction-tuned model in the Gemma 2 family, using the same sliding window + full attention hybrid and logit soft-capping as the 9B variant but at 2.6 billion parameters. At release it set a new bar for sub-3B instruction models on standard benchmarks. It is Apache 2.0 licensed and runs on consumer hardware.
by google
Gemma 2 9B Instruct is Google's instruction-tuned 9B model from the Gemma 2 family, which introduced sliding window + full attention alternation and logit soft-capping for improved training stability. At release it outperformed Llama 3 8B on multiple benchmarks while remaining smaller, making it one of the most downloaded open instruction models in its size class. It is English-focused with some multilingual capability.
by HuggingFaceTB
SmolLM3-3B is HuggingFace's 3B instruction-tuned language model, the third generation of the SmolLM family targeting on-device and resource-constrained deployment. It is multilingual (English, French, Spanish, Italian, Portuguese, Chinese, Arabic, Russian) and achieves competitive instruction-following quality at the 3B parameter scale. Apache-2.0 licensing makes it a viable base for commercial on-device AI applications.
by meta-llama
LLaMA 2 7B Chat is Meta's 7B RLHF-aligned conversational model from 2023. While superseded by LLaMA 3 and later releases, it remains a well-understood reference model used for fine-tuning experiments, benchmarking, and educational purposes.
by EleutherAI
GPT-Neo-125M is EleutherAI's open recreation of the GPT-2 class of models, pre-trained on the Pile dataset as part of their open language model initiative. At 125M parameters it's a pedagogical and baseline research model rather than a practical text generator. MIT-licensed and available in multiple frameworks.
by tiiuae
Falcon-7B was TII UAE's 7B autoregressive language model released in 2023, trained on the RefinedWeb dataset derived from Common Crawl with aggressive deduplication and filtering. At release it matched GPT-3.5 on several benchmarks while being fully open-weight. Falcon-7B is a base model without instruction tuning; it is notable historically as an early high-quality openly-licensed 7B LLM.
by deepseek-ai
DeepSeek-R1-Distill-Qwen-1.5B distills DeepSeek-R1's chain-of-thought reasoning traces into a 1.5B Qwen2 model. The distillation process transfers structured thinking patterns rather than raw capability, producing a model that generates explicit reasoning steps before answers. MIT license makes it broadly usable.
by nvidia
Kimi-K2.6-NVFP4 is an NVIDIA-optimized FP4 quantization of Kimi-K2.6, produced with the ModelOpt toolkit for deployment on NVIDIA H100/H200 GPUs. FP4 weights reduce GPU memory footprint roughly 2x compared to BF16 while maintaining most of the original accuracy for conversational tasks. It is intended for inference on NVIDIA TensorRT-LLM or vLLM backends, not for further fine-tuning.
by QuantTrio
An AWQ 4-bit quantisation of DeepSeek V3.2, packaged for vLLM inference. AWQ (Activation-aware Weight Quantisation) identifies and preserves the most salient weights at higher precision, typically losing less perplexity than naive 4-bit approaches. This checkpoint lets teams run the large DeepSeek V3.2 on fewer GPUs than the BF16 original while retaining most benchmark performance.
by meta-llama
Meta's Llama 3 8B base model, pretrained on over 15 trillion tokens with an expanded 128K token vocabulary. It serves as the foundation for instruction-tuned and task-specific finetunes in the Llama 3 ecosystem.
by nvidia
Nemotron-Mini-4B-Instruct is a 4B-parameter instruction-tuned language model from NVIDIA built on the LLaMA-3 architecture, targeting on-device and edge deployment scenarios where larger models are impractical. It is described in arXiv:2407.14679 and trained using NVIDIA's NeMo framework with alignment techniques from arXiv:2402.16819. The model is English-only and optimized for chat and assistant-style tasks.
by LiquidAI
LFM2.5-1.2B-Instruct is Liquid AI's 1.2B instruction-tuned model using their Liquid Foundation Model architecture, which combines recurrent and attention mechanisms for improved long-context efficiency. Supports 9 languages and is positioned as an edge-friendly model from a non-transformer architecture lineage. License is listed as 'other' — check Liquid AI's terms.
by huihui-ai
An abliterated (safety-removed) version of Qwen2.5-72B-Instruct by huihui-ai, where refusal mechanisms have been removed using directional activation manipulation. This allows the model to respond to requests the original would decline. The abliteration technique is reversible but the resulting model lacks safety guardrails.
by Qwen
Qwen3-32B-FP8 is Alibaba's official FP8-quantized checkpoint of the Qwen3-32B instruction-tuned model, targeting Hopper (H100) GPU inference with FP8 tensor core support. FP8 quantization reduces memory by ~50% vs bf16 while preserving most of the model's accuracy. Apache-2.0 licensed.
by Qwen
Qwen2.5-Coder-7B-Instruct in AWQ 4-bit quantisation, the official Alibaba release for memory-efficient code generation serving. AWQ preserves the most salient weights at higher precision, enabling deployment of the 7B code model on a single GPU with ~8GB VRAM. It achieves competitive HumanEval and MBPP scores relative to the BF16 original while halving memory requirements.
by Qwen
QwQ-32B is Alibaba's 32B-parameter reasoning-focused language model, designed to 'think slowly' through complex problems with long chain-of-thought generation before producing answers. With 2,950 likes it is one of the most popular open-weight reasoning models on HuggingFace.
by Qwen
Qwen3 80B MoE instruct model activating 3B parameters per token, offering a high-capacity but compute-efficient inference profile. Positioned as a next-generation step-up from the Qwen3-30B-A3B series with additional pretraining compute.
by facebook
OPT-1.3B is Meta's Open Pre-trained Transformer at 1.3 billion parameters, released in 2022 as part of a suite ranging from 125M to 175B. The model was trained on a curated mix of publicly available datasets and released with full weights and training logs to enable reproducibility research. It has largely been superseded by later open LLMs but remains a useful controlled baseline.
by LSX-UniWue
LLaMmlein 1B is a German-centric small language model from the University of Würzburg's LSX group, trained from scratch on German text. The 'prerelease' indicates this is a preliminary checkpoint shared before the final publication.
by janhq
Jan v3.5-4B is Homebrew (Jan.ai)'s 4B instruction-tuned model in GGUF format, designed for local deployment via the Jan desktop application and llama.cpp. It is fine-tuned for general assistant tasks including math, coding, and identity-aware conversation. Jan.ai positions this as a private, on-device alternative to cloud AI assistants for consumer use.
by stelterlab
An AWQ (Activation-aware Weight Quantization) conversion of Mistral Small 24B Instruct (January 2025), offering 4-bit quantized inference at reduced memory while preserving most of the original model's instruction-following quality.
by lmstudio-community
A 5-bit MLX quantization of LFM2-24B-A2B, sitting between the 4-bit and 8-bit variants in the accuracy/memory tradeoff space. Useful for Apple Silicon users who want more quality than 4-bit but less memory usage than 8-bit.
by google
Gemma 2B is Google's 2B-parameter open language model from early 2024, trained on 2T tokens of web, code, and math data. It was notable at release for punching above its weight class on benchmarks vs other 2B models available at the time.