# Can I Run This LLM?

> Local LLM memory calculator. Pick an open-weight model, quantization and context length to see the estimated VRAM or unified memory needed, and the cheapest GPU, Mac or AI mini PC that fits, priced from real offers tracked by AI Hardware Deals (https://hardware.branolabs.com). 42 models, data from Hugging Face, updated 2026-10-17.

Machine-readable data: https://canirun.branolabs.com/api/models.json

## Methodology
- Weights = total parameters (Hugging Face safetensors metadata) x bits per weight (llama.cpp quantize README) / 8.
- KV cache from config.json: full-attention layers x 2 x KV heads x head dim x bytes x tokens; sliding-window, hybrid and MLA layouts handled separately.
- Runtime overhead estimate: 1 GB + 3% of weights. All numbers are estimates.

## Models (estimated memory at Q4_K_M, 32K context)
- [Gemma 4 E2B](https://canirun.branolabs.com/model/gemma-4-e2b): 5.1B params, about 4.5 GB. Runs in: Ollama, LM Studio, llama.cpp, MLX (mlx-lm). Source: https://huggingface.co/google/gemma-4-E2B-it
- [Ling-3.0-tiny 7.9B-A1.3B](https://canirun.branolabs.com/model/ling-3-0-tiny): 7.9B params, 1.3B active, about 6.2 GB. Runs in: llama.cpp, vLLM, SGLang. Source: https://huggingface.co/inclusionAI/Ling-3.0-tiny
- [Gemma 4 E4B](https://canirun.branolabs.com/model/gemma-4-e4b): 8.0B params, about 6.5 GB. Runs in: Ollama, LM Studio, llama.cpp, MLX (mlx-lm). Source: https://huggingface.co/google/gemma-4-E4B-it
- [Llama 3.1 8B](https://canirun.branolabs.com/model/llama-3-1-8b): 8.0B params, about 10.4 GB. Runs in: Ollama, LM Studio, llama.cpp, MLX (mlx-lm). Source: https://huggingface.co/meta-llama/Llama-3.1-8B-Instruct
- [Qwen3 8B](https://canirun.branolabs.com/model/qwen3-8b): 8.2B params, about 11.0 GB. Runs in: Ollama, llama.cpp, MLX (mlx-lm), KTransformers, vLLM, SGLang. Source: https://huggingface.co/Qwen/Qwen3-8B
- [Qwen3.5 9B](https://canirun.branolabs.com/model/qwen3-5-9b): 9.7B params, about 8.2 GB. Runs in: Ollama, LM Studio, llama.cpp, MLX (mlx-lm), KTransformers, vLLM, SGLang. Source: https://huggingface.co/Qwen/Qwen3.5-9B
- [Gemma 4 12B](https://canirun.branolabs.com/model/gemma-4-12b): 12B params, about 11.0 GB. Runs in: Ollama, LM Studio, llama.cpp, MLX (mlx-lm). Source: https://huggingface.co/google/gemma-4-12B-it
- [Phi-4 14B](https://canirun.branolabs.com/model/phi-4): 15B params, about 16.9 GB. Runs in: Ollama, LM Studio, llama.cpp, MLX (mlx-lm). Source: https://huggingface.co/microsoft/phi-4
- [gpt-oss-20b](https://canirun.branolabs.com/model/gpt-oss-20b): 21B params, 3.6B active, about 15.0 GB. Runs in: Ollama, LM Studio, llama.cpp, MLX (mlx-lm), vLLM. Source: https://huggingface.co/openai/gpt-oss-20b
- [LFM2.5 24B-A2B](https://canirun.branolabs.com/model/lfm2-5-24b-a2b): 24B params, 2B active, about 16.8 GB. Runs in: Ollama, llama.cpp, MLX (mlx-lm), vLLM, SGLang. Source: https://huggingface.co/LiquidAI/LFM2.5-24B-A2B
- [Devstral Small 2 24B](https://canirun.branolabs.com/model/devstral-small-2-24b): 24B params, about 21.5 GB. Runs in: Ollama, LM Studio, llama.cpp, MLX (mlx-lm), vLLM, SGLang. Source: https://huggingface.co/mistralai/Devstral-Small-2-24B-Instruct-2512
- [Gemma 4 26B-A4B](https://canirun.branolabs.com/model/gemma-4-26b-a4b): 26B params, 4B active, about 18.8 GB. Runs in: Ollama, LM Studio, llama.cpp, MLX (mlx-lm). Source: https://huggingface.co/google/gemma-4-26B-A4B-it
- [Qwen3.8 27B](https://canirun.branolabs.com/model/qwen3-8-27b): 28B params, about 20.7 GB. Runs in: colibri, Ollama, LM Studio, llama.cpp, MLX (mlx-lm), vLLM, SGLang, TokenSpeed. Source: https://huggingface.co/Qwen/Qwen3.8-27B
- [Qwen3.6 27B](https://canirun.branolabs.com/model/qwen3-6-27b): 28B params, about 20.7 GB. Runs in: Ollama, LM Studio, llama.cpp, MLX (mlx-lm), KTransformers, vLLM, SGLang. Source: https://huggingface.co/Qwen/Qwen3.6-27B
- [Qwen3-Coder 30B-A3B](https://canirun.branolabs.com/model/qwen3-coder-30b-a3b): 31B params, 3.3B active, about 23.5 GB. Runs in: colibri, Ollama, LM Studio, llama.cpp, MLX (mlx-lm), KTransformers. Source: https://huggingface.co/Qwen/Qwen3-Coder-30B-A3B-Instruct
- [GLM-4.7-Flash 30B-A3B](https://canirun.branolabs.com/model/glm-4-7-flash): 31B params, 3B active, about 22.4 GB. Runs in: Ollama, LM Studio, llama.cpp, MLX (mlx-lm), vLLM, SGLang. Source: https://huggingface.co/zai-org/GLM-4.7-Flash
- [Gemma 4 31B](https://canirun.branolabs.com/model/gemma-4-31b): 31B params, about 26.9 GB. Runs in: Ollama, LM Studio, llama.cpp, MLX (mlx-lm). Source: https://huggingface.co/google/gemma-4-31B-it
- [Nemotron 3 Nano 30B-A3B](https://canirun.branolabs.com/model/nemotron-3-nano-30b-a3b): 32B params, 3B active, about 21.1 GB. Runs in: Ollama, LM Studio, llama.cpp, MLX (mlx-lm), vLLM, SGLang. Source: https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16
- [Qwen3 32B](https://canirun.branolabs.com/model/qwen3-32b): 33B params, about 30.2 GB. Runs in: Ollama, LM Studio, llama.cpp, MLX (mlx-lm), KTransformers, vLLM, SGLang. Source: https://huggingface.co/Qwen/Qwen3-32B
- [Qwen3.6 35B-A3B](https://canirun.branolabs.com/model/qwen3-6-35b-a3b): 36B params, 3B active, about 24.3 GB. Runs in: Strata, colibri, Ollama, LM Studio, llama.cpp, MLX (mlx-lm), KTransformers, vLLM, SGLang. Source: https://huggingface.co/Qwen/Qwen3.6-35B-A3B
- [Llama 3.3 70B](https://canirun.branolabs.com/model/llama-3-3-70b): 71B params, about 56.2 GB. Runs in: Ollama, LM Studio, llama.cpp, MLX (mlx-lm), vLLM. Source: https://huggingface.co/meta-llama/Llama-3.3-70B-Instruct
- [Llama 4 Scout 17B-16E](https://canirun.branolabs.com/model/llama-4-scout): 109B params, 17B active, about 75.9 GB. Runs in: Ollama, LM Studio, llama.cpp, MLX (mlx-lm). Source: https://huggingface.co/meta-llama/Llama-4-Scout-17B-16E-Instruct
- [GLM-4.5-Air](https://canirun.branolabs.com/model/glm-4-5-air): 110B params, 12B active, about 76.8 GB. Runs in: LM Studio, llama.cpp, MLX (mlx-lm), vLLM, SGLang. Source: https://huggingface.co/zai-org/GLM-4.5-Air
- [gpt-oss-120b](https://canirun.branolabs.com/model/gpt-oss-120b): 117B params, 5.1B active, about 75.8 GB. Runs in: Ollama, LM Studio, llama.cpp, MLX (mlx-lm), vLLM. Source: https://huggingface.co/openai/gpt-oss-120b
- [Mistral Small 4 119B](https://canirun.branolabs.com/model/mistral-small-4-119b): 119B params, 6.5B active, about 77.0 GB. Runs in: LM Studio, llama.cpp, MLX (mlx-lm), vLLM, SGLang. Source: https://huggingface.co/mistralai/Mistral-Small-4-119B-2603
- [Nemotron 3 Super 120B-A12B](https://canirun.branolabs.com/model/nemotron-3-super-120b-a12b): 124B params, 12B active, about 79.2 GB. Runs in: Ollama, LM Studio, llama.cpp, MLX (mlx-lm), vLLM, SGLang. Source: https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-BF16
- [Qwen3.5 122B-A10B](https://canirun.branolabs.com/model/qwen3-5-122b-a10b): 125B params, 10B active, about 80.6 GB. Runs in: Ollama, LM Studio, llama.cpp, MLX (mlx-lm), KTransformers, vLLM, SGLang. Source: https://huggingface.co/Qwen/Qwen3.5-122B-A10B
- [Ling-3.0-flash 124B-A5.1B](https://canirun.branolabs.com/model/ling-3-0-flash): 127B params, 5.1B active, about 81.6 GB. Runs in: llama.cpp, vLLM, SGLang. Source: https://huggingface.co/inclusionAI/Ling-3.0-flash
- [Mistral Medium 3.5 128B](https://canirun.branolabs.com/model/mistral-medium-3-5-128b): 128B params, about 93.3 GB. Runs in: Ollama, llama.cpp, MLX (mlx-lm), vLLM, SGLang. Source: https://huggingface.co/mistralai/Mistral-Medium-3.5-128B
- [Qwen3.8-Flash-Next 125B-A6B](https://canirun.branolabs.com/model/qwen3-8-flash-next): 180B params, 6B active, about 115 GB. Runs in: Strata, colibri, Ollama, LM Studio, llama.cpp, MLX (mlx-lm), KTransformers, vLLM, SGLang, TokenSpeed. Source: https://huggingface.co/Qwen/Qwen3.8-Flash-Next
- [Step 3.5 Flash 196B-A11B](https://canirun.branolabs.com/model/step-3-5-flash): 199B params, about 128 GB. Runs in: llama.cpp, MLX (mlx-lm), vLLM, SGLang. Source: https://huggingface.co/stepfun-ai/Step-3.5-Flash
- [MiniMax M2.7](https://canirun.branolabs.com/model/minimax-m2-7): 229B params, about 153 GB. Runs in: LM Studio, llama.cpp, MLX (mlx-lm), vLLM, SGLang. Source: https://huggingface.co/MiniMaxAI/MiniMax-M2.7
- [DeepSeek V4 Flash (0731)](https://canirun.branolabs.com/model/deepseek-v4-flash): 304B params, 13B active, about 193 GB. Runs in: colibri, LM Studio, llama.cpp, vLLM, SGLang. Source: https://huggingface.co/deepseek-ai/DeepSeek-V4-Flash-0731
- [MiMo-V2.6-Flash 309B-A15B](https://canirun.branolabs.com/model/mimo-v2-6-flash): 311B params, 15B active, about 198 GB. Runs in: colibri, llama.cpp, vLLM, SGLang. Source: https://huggingface.co/XiaomiMiMo/MiMo-V2.6-Flash-MOPD
- [GLM-5.3-Flash 320B-A18B](https://canirun.branolabs.com/model/glm-5-3-flash): 321B params, 18B active, about 204 GB. Runs in: colibri, llama.cpp, MLX (mlx-lm), KTransformers, vLLM, SGLang, TokenSpeed. Source: https://huggingface.co/zai-org/GLM-5.3-Flash
- [Qwen3-Coder 480B-A35B](https://canirun.branolabs.com/model/qwen3-coder-480b-a35b): 480B params, 35B active, about 312 GB. Runs in: Ollama, LM Studio, llama.cpp, MLX (mlx-lm), KTransformers. Source: https://huggingface.co/Qwen/Qwen3-Coder-480B-A35B-Instruct
- [DeepSeek V3.1](https://canirun.branolabs.com/model/deepseek-v3-1): 685B params, 37B active, about 435 GB. Runs in: Ollama, llama.cpp, MLX (mlx-lm). Source: https://huggingface.co/deepseek-ai/DeepSeek-V3.1
- [GLM-5.3](https://canirun.branolabs.com/model/glm-5-3): 753B params, about 479 GB. Runs in: colibri, llama.cpp, MLX (mlx-lm), KTransformers, vLLM, SGLang, TokenSpeed. Source: https://huggingface.co/zai-org/GLM-5.3
- [DeepSeek V4.1 Flash](https://canirun.branolabs.com/model/deepseek-v4-1-flash): 763B params, 16B active, about 482 GB. Runs in: colibri. Source: https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash
- [Kimi K2 Instruct](https://canirun.branolabs.com/model/kimi-k2): 1.03T params, 32B active, about 650 GB. Runs in: llama.cpp, MLX (mlx-lm), KTransformers, vLLM, SGLang. Source: https://huggingface.co/moonshotai/Kimi-K2-Instruct
- [DeepSeek V4 Pro](https://canirun.branolabs.com/model/deepseek-v4-pro): 1.60T params, 49B active, about 1009 GB. Runs in: MLX (mlx-lm). Source: https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro
- [Kimi K3](https://canirun.branolabs.com/model/kimi-k3): 2.78T params, 104B active, about 1754 GB. Runs in: colibri, llama.cpp (fork only), vLLM, SGLang, TokenSpeed. Source: https://huggingface.co/moonshotai/Kimi-K3

## Local apps and engines (https://canirun.branolabs.com/tools)
- [Strata](https://github.com/Niko1221/Strata): Single-model engine + one-click installer. Windows 10/11, Linux, macOS (Apple Silicon, experimental). Formats: GGUF quants of Qwen3.8-Flash-Next (ISTA-DASLab GSQ-RCO, Unsloth, UkisAI Swift 1.5); uses parts of llama.cpp/ggml. License: MIT.
- [colibri](https://github.com/JustVugg/colibri): Disk-streaming MoE engine. Linux, Windows, macOS. Formats: Its own int4/int8 containers, converted once from the Hugging Face safetensors (pre-converted copies on Hugging Face). License: Apache-2.0.
- [Ollama](https://ollama.com): Desktop app + CLI. macOS, Windows, Linux. Formats: Models from the Ollama library (GGUF-based; MLX builds on Mac for some models). License: MIT.
- [LM Studio](https://lmstudio.ai): Desktop app. macOS, Windows (x64, arm64), Linux (x64, arm64). Formats: GGUF and MLX. License: Free app (closed source); lms CLI is MIT.
- [llama.cpp](https://github.com/ggml-org/llama.cpp): Inference engine (C/C++). Linux, Windows, macOS, Android (build). Formats: GGUF. License: MIT.
- [MLX (mlx-lm)](https://github.com/ml-explore/mlx-lm): Apple framework + CLI. macOS (Apple Silicon only). Formats: MLX safetensors (4/6/8-bit, mxfp4). License: MIT.
- [KTransformers](https://github.com/kvcache-ai/ktransformers): CPU+GPU hybrid engine. Linux, Windows. Formats: Hugging Face weights, CPU-side INT4/INT8, GPTQ. License: Apache-2.0.
- [vLLM](https://vllm.ai): Serving engine. Linux. Formats: Hugging Face safetensors, FP8, AWQ/GPTQ, NVFP4. License: Apache-2.0.
- [SGLang](https://sglang.io): Serving engine. Linux. Formats: Hugging Face safetensors, FP8 and more. License: Apache-2.0.
- [TokenSpeed](https://lightseek.org/blog/lightseek-tokenspeed.html): Serving engine. Linux. Formats: Hugging Face weights, FP8/FP4. License: MIT.
- [Jan](https://jan.ai): Desktop app. Windows, macOS, Linux. Formats: GGUF. License: Open source (see repo).
- [PocketPal AI](https://github.com/a-ghorbani/pocketpal-ai): Phone app. iOS / iPadOS, Android. Formats: GGUF. License: MIT.

## API only, no local weights
- Qwen3.8-Flash (Qwen (Qwen Cloud)): The production API version built on Qwen3.8-Flash-Next, with 1M context by default and built-in tools. Run Qwen3.8-Flash-Next locally instead. https://www.qwencloud.com/models/qwen3.8-flash

## Optional
- [Methodology](https://canirun.branolabs.com/about)
- [All models](https://canirun.branolabs.com/models)
- [Local apps](https://canirun.branolabs.com/tools)
- [AI Hardware Deals](https://hardware.branolabs.com): the price source
- [@BranoLabs on X](https://x.com/BranoLabs): updates and feedback