Local AI model library

DeepSeek-R1 distilled to Llama 70B base. Strong general reasoning.

DeepSeek-R1 reasoning model distilled to 32B. Excellent at math & logic.

DeepSeek V4.1 Flash is DeepSeek's MIT-licensed open-weight multimodal MoE for reasoning, coding, agents, and native image understanding. The complete checkpoint is approximately 763B parameters, including a 552B backbone; its Causal Encoder-Decoder architecture activates 8B parameters during input prefill and 16B during output decoding. CSA2 and FP4 KV caching reduce long-context cache cost. The model supports a 1M-token context window and is available through official FP8 weights, the DeepSeek API, and Ollama Cloud. It remains a data-center-class deployment despite the low active-parameter count.

DeepSeek V4.1 Flash 1017 is an updated MIT-licensed 763B-parameter multimodal MoE checkpoint. It keeps the V4.1 family's 1M-token context, coding, reasoning, tool use, and image understanding, while remaining a data-center-class deployment despite activating far fewer parameters per token.

Alibaba's flagship Qwen3. Competitive with GPT-4 class models.

Microsoft's compact 14B. Punches way above its weight, especially on math.

Z.ai's GLM 5.2 FP8.

Z.ai's GLM 5.1 FP8.

Z.ai's GLM 4.7 Flash.

Moonshot AI's Kimi K2.7 Code.

Moonshot AI's Kimi K2.6.

Moonshot AI's Kimi K2 Thinking.

MiniMax's MiniMax M3.

MiniMax's MiniMax M3 MXFP8.

MiniMax's MiniMax M2.7.

Google's official instruction-tuned Gemma 4 31B checkpoint with W4A16 quantization-aware compressed weights for text and image understanding. The published repository is downloadable; RAM and VRAM figures on this page remain runtime planning estimates.

Google's Gemma 4 12B It QAT W4a16 Ct.

Google's Gemma 4 E4B It QAT Q4 0 GGUF.

Black Forest Labs's FLUX.2 Dev.

Black Forest Labs's FLUX.2 Klein 9B.

Black Forest Labs's FLUX.2 Klein 4B.

Alibaba's Wan2.2 T2V A14B.

Alibaba's Wan2.2 TI2V 5B.

Alibaba's Wan2.2 I2V A14B.

Lightricks's LTX 2.3.

Lightricks's LTX 2.3 FP8.

Z.ai's GLM 5.2.

Moonshot AI Kimi K3 is an open-weight 2.8T-parameter multimodal MoE model with 104B active parameters, native MXFP4 weights, always-on reasoning, tool use, and a 1M-token context window. The official release targets large multi-accelerator deployments through vLLM, SGLang, or TokenSpeed rather than consumer local hardware.

Qwen3.8-2.4T-A95B is Alibaba Qwen's open-weight, text-only flagship MoE model with 2.4T total parameters and 95B activated parameters. It uses mandatory thinking, supports tool use, has a native 262K context window extensible to about 1.01M tokens, and ships in official BF16 and FP8 repositories. Its multi-terabyte weights target distributed data-center inference with vLLM, SGLang, or TokenSpeed rather than consumer local hardware.

DeepSeek V4 Flash 0731 is the official open-weight release that superseded the preview model and added the DSpark speculative decoding module. Its 304B-parameter, approximately 13B-active MoE checkpoint uses mixed FP4/FP8 weights, supports a 1M-token context window, three reasoning-effort levels, and agentic tool use. The downloadable weights remain available, but DeepSeek retired the hosted V4 Flash API alias in September 2026 in favor of DeepSeek V4.1 Flash.

Meta Muse Glimmer 30B is an Apache 2.0 open-weight, dense multimodal agentic model designed for always-on local workflows. It combines text and image understanding, long-context reasoning, coding, tool calling, and computer-use capabilities. Meta publishes BF16 weights, official GGUF quantizations targeting 24GB and 32GB VRAM, and ExecuTorch packages for Metal and CUDA.

Qwen3.8-27B is Alibaba Qwen's Apache 2.0 open-weight dense vision-language model for coding, professional work, research, and long-horizon agents. It has 27B parameters, a native 262K context window extensible to 1M tokens, flexible reasoning effort, tool use, official BF16 and FP8 weights, and a broad community quantization ecosystem for consumer GPUs and Apple Silicon.

Qwen3.8-Flash-Next is Alibaba Qwen's open-weight multimodal MoE preview of the architecture planned for Qwen4. Its language model has 125B parameters with 6B activated per token, plus 51B n-gram embeddings and 4B MTP parameters. It combines Gated DeltaNet, Qwen Sparse Attention, gated residuals, vision input, tool use, flexible reasoning, a native 262K context window extensible to 1M tokens, and official serving support through vLLM, SGLang, KTransformers, and TokenSpeed.

GLM-5.3-Flash is Z.ai's MIT-licensed, open-weight multimodal MoE for coding, long-horizon agents, computer use, and professional work. It has 320B total parameters with 18B activated per token, hybrid linear and sparse attention, native image input, a 1M-token context window, and official support for SGLang, vLLM, and TokenSpeed. Its total weight footprint makes it a server-class local model despite its efficient active-parameter count.

LTX-2.5 is Lightricks' 22B joint audio-video generation model for text-to-video, image-to-video, video editing, and synchronized audio generation. It ships through official single-file, Diffusers, and ComfyUI workflows under the LTX 2.x Community License.

Alibaba's open-weight Qwen3.6 27B. Strong coding-agent, reasoning, long-context, and vision-language model with 262K native context.

Meta Muse Spark Open is a 816B-total, 42B-active multimodal MoE with a shared 1M-token context across text, image, video, and audio inputs. It is designed for coding and agentic work under the custom Muse Model License; the base checkpoint is a distributed server-class model, with a separate official NVFP4 release.

Qwen-Image 2.1 Turbo is the accelerated 7B checkpoint of Qwen-Image 2.1 for text-to-image generation and image editing. It uses eight denoising steps, CFG 1 by default, and prefix KV caching, and is released under the Qwen Research License.

NVIDIA Nemotron 3.5 Super VL is a 120B-total, 12B-active vision-language MoE for reasoning, coding, tool use, document understanding, and video understanding. This official NVFP4 checkpoint supports up to a 1M-token context and targets multi-GPU NVIDIA deployments.

Alibaba's Qwen3-Coder 30B with MoE architecture. Top coding model in <30B class.

Reflection Beam 1.0 is an Apache 2.0 open-weight 501B-parameter MoE with 23B active parameters and a 1M-token context window. It targets reasoning, coding, agents, and tool use; the base BF16 checkpoint requires distributed server hardware, with separate official FP8 and NVFP4 repositories for reduced-memory deployment.

Black Forest Labs FLUX.3 Image Klein is a 5.6B open-weight model for text-to-image generation and image editing. The distilled checkpoint targets 8-step generation and includes Diffusers, ComfyUI, and repository-hosted NVFP4 variants under the FLUX Kommunity License.

Qwen 2.5 Coder 32B. Predecessor to Qwen3-Coder, still very capable.

MiniCPM-V 4.7 35B-A3B is OpenBMB's larger Apache 2.0 multimodal MoE, with about 35B total and 3B active parameters. It targets stronger image, document, and video understanding while retaining official GGUF, FP8, AWQ, GPTQ, and INT4 deployment options.

Moonshot AI's Kimi K2.5.

MiniMax's MiniMax M2.5.

MiniMax's MiniMax M2.

DeepSeek's efficient MoE coder. 16B total / 2.4B active.

Meta's Llama 3.3 70B. Strong all-rounder, 128K context.

Moonshot AI's Kimi K2 Instruct 0905.

Google's Diffusiongemma 26B A4B It.

Black Forest Labs's FLUX.1 Kontext Dev.

Stability AI's Stable Diffusion 3.5 Medium.

Bespoke Nimble 9B is an Apache-2.0 PEFT/LoRA adapter for Qwen3.5-9B, not a standalone 9B checkpoint. It scores enum, boolean, and rubric answers directly, supports up to 255 choices per field and an 8K request limit, and requires the separate Qwen3.5-9B base weights. Its narrow author-reported evaluation should be recalibrated on the target workflow.

Liquid AI LFM2.5 24B-A2B is a 24B-total, 2B-active reasoning MoE built for fast private agents on-device. It provides a 131K-token context, improved multilingual tokenization, tool-oriented post-training, and official native and GGUF checkpoints.

Mistral's coding-focused model. Strong agent capabilities.

Moonshot AI's Kimi VL A3B Thinking 2506.

Stability AI's Stable Diffusion 3.5 Large.

Alibaba's Wan2.2 Animate 14B.

Qwen3 30B with MoE. Same family as Coder, general-purpose.

Alibaba's Wan2.2 S2V 14B.

Lightricks's LTX 2.3 Nvfp4.

Laya Typed Decisions is the 421M-parameter checkpoint specialised for choice, ordinal score, and yes/no workflows. It is the strongest default Laya checkpoint for English typed-decision applications, while the built-in Router can select another family member by language and request shape. Use a labelled golden set and explicit human-review thresholds before allowing high-impact actions.

Liquid AI d1-3B is a compact multimodal decision model built on LFM2.5-VL-3B. It produces calibrated choices for bounded visual and text decisions with a 32K-token context; it is not a general chat or coding model and should be validated on a labelled task-specific set.

Mistral Small 3.1 with vision. Strong multilingual support.

Google's Gemma 4 E2B It QAT Mobile Transformers.

MiniCPM-V 4.7 1B is OpenBMB's compact Apache 2.0 multimodal model for image, document, and video understanding. Official GGUF and low-bit variants make it the local-friendly member of the 4.7 family.

DeepSeek-VL2 is the full 27.5B-parameter MoE vision-language model, activating about 4.5B parameters per token. It provides the strongest capability in the VL2 family for visual question answering, OCR, document, table and chart understanding, and visual grounding.

DeepSeek-OCR 2 is DeepSeek's 3B-parameter document understanding model based on Visual Causal Flow. It converts document images into structured text or Markdown and supports dynamic-resolution OCR, layout grounding, and local serving through Transformers, vLLM, and SGLang.

Laya is the 421M-parameter English checkpoint in Convai Innovations' open System One family. Its bidirectional ModernBERT encoder answers typed choice, score, and yes/no questions in one non-autoregressive forward pass, returning probabilities instead of generated text. It is intended for bounded routing, triage, guardrails, moderation, and workflow decisions and should be evaluated on each application's own labelled examples before automation.

DeepSeek-VL2 Small is the middle 16.1B-parameter MoE member of DeepSeek's vision-language family, activating about 2.8B parameters per token. It is designed for visual question answering, OCR, document, table and chart understanding, and visual grounding.

Laya Multilingual is the 322M-parameter mmBERT member of the Laya family for typed decisions across more than 100 languages. The router uses it for non-English state and it supports longer inputs than the English checkpoint, with a 1K normal operating context and an advertised maximum up to 8K. Confidence still requires workload-specific validation and must not be treated as a guarantee.

Janus-Pro 7B is DeepSeek's larger unified multimodal model for image understanding and text-to-image generation. It improves instruction following and generation stability while retaining separate visual encoding paths for understanding and generation.

Google's open multimodal 27B. Vision + language in one model.

Qwen3 8B. Compact general-purpose, runs on any laptop.

DeepSeek-VL2 Tiny is the smallest official DeepSeek-VL2 vision-language model. Its 3B-parameter MoE activates about 1B parameters per token and targets visual question answering, OCR, document and chart understanding, and visual grounding with a 4K context window.

Qwen 2.5 VL (Vision-Language). Top open vision model.

LLaVA 1.6 multimodal model. Image understanding + chat.

Alibaba's Wan 2.1 text-to-video. Open weights, runs on consumer GPUs.

Zhipu AI's CogVideoX 5B. Compact open video model.

BAAI's BGE-M3. Multilingual embedding for RAG.

Nomic AI's text embedding. Long context, fully open.

OpenAI's Whisper Large v3. Top open ASR model.

F5-TTS. Open-source TTS with voice cloning.

Alibaba's CosyVoice. Multilingual TTS with emotion control.

Hugging Face's SmolLM2 1.7B. Tiny but capable, fits anywhere.

Janus-Pro 1B is DeepSeek's compact unified multimodal model for image understanding and text-to-image generation. Its decoupled visual encoding paths separate understanding from generation while sharing one transformer architecture.

Black Forest Labs' state-of-the-art text-to-image model. Top of GenEval.

Stability AI's classic SDXL. Best bang/buck for image gen.

Tencent's open HunyuanVideo. 13B params, high quality.

Meta's tiny Llama 3.2 3B. Runs on phones & weak hardware.

Google's Gemma 3n. Multimodal (text/vision/audio) at 4B size.

Stable Diffusion 1.5 is a mature text-to-image baseline with broad ComfyUI support and a relatively modest local memory requirement.