# The AI Bench

> Checks which AI models run on your own computer (graphics card, Mac or CPU), how fast they will run, and how to start them. Free, no sign-up. Data checked 10 October 2026.

Memory needed = model file + conversation memory (KV cache) + about 1 GB. Speed is estimated from memory bandwidth with constants fitted on public llama.cpp and MLX measurements; see the method page.

## Tool and API
- [Checker](https://theaibench.ai/): pick a graphics card or Mac, a task and a conversation length.
- [API and CLI](https://theaibench.ai/api/): GET https://theaibench.ai/api/v1/plan?platform=windows&vram=24&use_case=coding returns JSON. CLI: uvx theaibench.
- [Method and about](https://theaibench.ai/about/)
- [Measured speeds](https://theaibench.ai/bench/): community llama.cpp and MLX runs with sources, and how close our estimate is.

## Guides
- [How much memory do you need to run AI at home?](https://theaibench.ai/guides/how-much-memory-do-i-need/): A plain guide to graphics memory for local AI: what uses it, how big each model is, and what 8, 12, 16, 24, 32, 64 and 128 GB can run.
- [Ollama vs LM Studio vs Jan: which app should you use?](https://theaibench.ai/guides/ollama-vs-lm-studio-vs-jan/): Three free apps run AI models on your own computer. Which one to pick, in plain English: Ollama for commands and other apps, LM Studio for a Mac-friendly window, Jan for fully open source.
- [Mac or PC for local AI? Capacity vs speed, explained](https://theaibench.ai/guides/mac-vs-pc-for-local-ai/): A Mac holds bigger AI models; a PC graphics card runs smaller ones faster. Plain-English comparison with real numbers from our checker.
- [When not to run AI locally (and when you should)](https://theaibench.ai/guides/when-not-to-run-local-ai/): Local AI is private and free to use, but cloud AI is still better at some jobs. A plain list of when to stay with ChatGPT or Claude, and when local wins.
- [AMD Radeon for local AI: what works in 2026](https://theaibench.ai/guides/amd-radeon-for-local-ai-rocm/): Running local AI on an AMD Radeon card: Vulkan or ROCm, which cards are worth it, and what to expect compared to NVIDIA. Plain-English guide.

## Hardware
- [NVIDIA RTX 5090](https://theaibench.ai/hardware/rtx-5090/): 32 GB. The fastest single card you can buy, at a very high 2026 price.
- [NVIDIA RTX 4090](https://theaibench.ai/hardware/rtx-4090/): 24 GB. Still a top 24 GB card. Runs the best 27–32B models at a smooth speed.
- [NVIDIA RTX 3090 (used)](https://theaibench.ai/hardware/rtx-3090-used/): 24 GB. The best value for 24 GB. Six years old, still the community favourite.
- [Two RTX 3090s (used)](https://theaibench.ai/hardware/dual-rtx-3090/): 48 GB. 48 GB of graphics memory for less than one new 32 GB card. Loud and power-hungry.
- [NVIDIA RTX 5080](https://theaibench.ai/hardware/rtx-5080/): 16 GB. Fast, but only 16 GB. Great for models up to about 20B.
- [NVIDIA RTX 5070 Ti](https://theaibench.ai/hardware/rtx-5070-ti/): 16 GB. Nearly the RTX 5080’s speed for less money. Same 16 GB limit.
- [NVIDIA RTX 5060 Ti 16 GB](https://theaibench.ai/hardware/rtx-5060-ti-16gb/): 16 GB. NVIDIA’s cheapest new card with 16 GB. Quiet and low power, but slower.
- [NVIDIA RTX 3060 12 GB](https://theaibench.ai/hardware/rtx-3060-12gb/): 12 GB. The cheapest useful NVIDIA card. Fine for 8–12B models.
- [AMD Radeon RX 9070 XT](https://theaibench.ai/hardware/amd-rx-9070-xt/): 16 GB. AMD’s fastest current 16 GB card. Good speed; software takes a little more care.
- [AMD Radeon RX 7900 XTX](https://theaibench.ai/hardware/amd-rx-7900-xtx/): 24 GB. 24 GB with RTX 4090-class bandwidth. Best if you already own one.
- [Intel Arc B580](https://theaibench.ai/hardware/intel-arc-b580/): 12 GB. Cheap 12 GB card. Works for small models, with more setup.
- [NVIDIA RTX A6000 (used)](https://theaibench.ai/hardware/rtx-a6000-used/): 48 GB. 48 GB on one card. Slower than new cards, but fits 70B models.
- [Mac mini (M6, 16 GB)](https://theaibench.ai/hardware/mac-mini-m6-16gb/): 16 GB. A quiet always-on box for small models (up to about 8B).
- [Mac mini (M5 Pro, 24 GB)](https://theaibench.ai/hardware/mac-mini-m5-pro-24gb/): 24 GB. A quiet desk machine for models up to about 14B, and faster than the M6.
- [MacBook Air (M5, 24 GB)](https://theaibench.ai/hardware/macbook-air-m5-24gb/): 24 GB. Silent and portable. Fine for small models; slow on big ones.
- [MacBook Pro (M5 Pro, 48 GB)](https://theaibench.ai/hardware/macbook-pro-m5-pro-48gb/): 48 GB. A laptop that runs 27–35B models. The sweet spot for developers on the move.
- [MacBook Pro (M5 Max, 64 GB)](https://theaibench.ai/hardware/m5-max-macbook-pro-64gb/): 64 GB. The fastest laptop for local AI. Runs 70B-class models.
- [Mac Studio (M5 Max, 64 GB)](https://theaibench.ai/hardware/mac-studio-m5-max-64gb/): 64 GB. A silent desk machine for 30–70B models. Good value among Macs.
- [Mac Studio (M5 Ultra, 96 GB)](https://theaibench.ai/hardware/mac-studio-m5-ultra-96gb/): 96 GB. The most capable Mac for local AI. Runs 70–120B models.
- [NVIDIA DGX Spark](https://theaibench.ai/hardware/nvidia-dgx-spark/): 128 GB. A small box with 128 GB that runs 100B+ mixture-of-experts models.
- [Framework Desktop (Ryzen AI Max+ 395, 128 GB)](https://theaibench.ai/hardware/framework-desktop-strix-halo/): 128 GB. A small PC with 128 GB shared memory. Great for big mixture-of-experts models.

## Text models
- [Qwen3.8 27B](https://theaibench.ai/models/qwen-3-8/): A leading 2026 all-round model that fits a 24 GB graphics card. Reads images and video too.
- [Qwen3.8 Flash Next](https://theaibench.ai/models/qwen-3-8-flash-next/): An experimental preview of the next Qwen design, and the most-used model on 128 GB machines. Smaller 3-bit and 4-bit files fit 96 GB.
- [Qwen3.6](https://theaibench.ai/models/qwen-3-6/): April 2026 Qwen focused on coding agents: a fast 35B-A3B and a 27B dense, both reading images.
- [Qwen3.5](https://theaibench.ai/models/qwen-3-5/): A full family from phone size (0.8B) to 122B, every size reading images. A strong pick for small machines.
- [Qwen3](https://theaibench.ai/models/qwen3/): The 2025 Qwen text models. Older now, but supported everywhere and well tested.
- [Qwen3-Coder](https://theaibench.ai/models/qwen3-coder/): Qwen models made for coding agents. The 30B-A3B fits a 24 GB card and answers fast.
- [Qwen3-VL](https://theaibench.ai/models/qwen3-vl/): Qwen’s 2025 vision models for photos, documents and screen control.
- [Gemma 4](https://theaibench.ai/models/gemma-4/): Google’s 2026 family, from phone size to 31B. Reads images and works in 140+ languages.
- [gpt-oss](https://theaibench.ai/models/gpt-oss/): OpenAI’s open models for reasoning and tool use. The 20B fits a 16 GB card and is fast.
- [Llama 3](https://theaibench.ai/models/llama-3/): Meta’s 2024 models. Older and weaker than 2026 models, but supported by every tool.
- [Muse Glimmer 30B](https://theaibench.ai/models/muse-glimmer/): Meta’s August 2026 open agent model, built to run tool-using tasks on one consumer graphics card.
- [Phi-4](https://theaibench.ai/models/phi-4/): Microsoft’s small models trained on textbook-style data. Good at maths for their size.
- [Ministral 3](https://theaibench.ai/models/ministral-3/): Mistral’s small models for laptops and 12–16 GB cards, with image input and a 256K context.
- [Devstral Small 2](https://theaibench.ai/models/devstral-small-2/): Mistral’s 24B coding-agent model. Mistral says it runs on one RTX 4090 or a 32 GB Mac.
- [Mistral Small 4](https://theaibench.ai/models/mistral-small-4/): A 119B mixture-of-experts model that merges Mistral’s chat, reasoning and coding lines. For 96–128 GB machines.
- [Mistral Medium 3.5](https://theaibench.ai/models/mistral-medium-3-5/): Mistral’s 128B dense flagship. Fits 96–128 GB machines, but every word uses all 128B parameters, so it is slow.
- [IBM Granite 4.2](https://theaibench.ai/models/granite-4/): IBM’s business-focused models (3B, 8B, 30B) with a thinking mode and tool calls.
- [NVIDIA Nemotron 3](https://theaibench.ai/models/nemotron-3/): NVIDIA reasoning models with a hybrid design that keeps conversation memory tiny, even at long context.
- [Nemotron 3.5 Lightning](https://theaibench.ai/models/nemotron-3-5-lightning/): NVIDIA’s August 2026 30B-A3B, released with its training data under a permissive licence.
- [GLM-4.7 Flash and GLM-4.5 Air](https://theaibench.ai/models/glm-4/): The Z.ai models you can run at home: a 30B-A3B for 24 GB machines and a 106B for 64–96 GB.
- [Ornith 1.5](https://theaibench.ai/models/ornith/): Coding-agent models built on Qwen3.5 and Gemma 4, in a 9B and a fast 35B-A3B.
- [Laguna 2.1](https://theaibench.ai/models/laguna/): Poolside’s coding-agent models: XS (33B-A3B) for 24–36 GB machines, S (118B) for 128 GB.
- [North Mini Code](https://theaibench.ai/models/north-mini-code/): Cohere’s 30B-A3B research model for code and terminal tasks.
- [MiMo-V2.6 9B](https://theaibench.ai/models/mimo-v2-6/): Xiaomi’s small local edition of MiMo-V2.6: Qwen3.5 9B trained further on agent and coding data.
- [MiniCPM5](https://theaibench.ai/models/minicpm5/): Tiny 1B and 2B models for phones and old laptops, with open training data.
- [LFM2.5](https://theaibench.ai/models/lfm-2-5/): Liquid AI’s small on-device models, tuned for tool use rather than coding or knowledge.
- [SmolLM3 3B](https://theaibench.ai/models/smollm3/): A fully open 3B model from Hugging Face: weights, data and training recipe all published.
- [Olmo 3](https://theaibench.ai/models/olmo-3/): Ai2’s fully open models (7B and 32B), each in a chat and a step-by-step thinking version.
- [Kimi Linear 48B-A3B](https://theaibench.ai/models/kimi-linear/): Moonshot’s long-context model: 1M tokens with a tiny conversation memory.
- [GLM-5.3-Flash](https://theaibench.ai/models/glm-5-3-flash/): Z.ai’s 320B model that reads images, with a 1M context. The main pick for 192 to 512 GB machines.
- [DeepSeek V4 Flash](https://theaibench.ai/models/deepseek-v4-flash/): DeepSeek’s 284B model with a 1M context and very small conversation memory. Fits 192 GB at 4-bit and 128 GB at 3-bit.

## Image models
- [Z-Image-Turbo](https://theaibench.ai/models/z-image-turbo/): A 6B image model that makes a picture in 8 steps and fits a 16 GB card. Apache 2.0.
- [FLUX.2](https://theaibench.ai/models/flux-2/): Black Forest Labs image models. The 4B “klein” is small, fast and free for business; the 9B and the big “dev” are non-commercial.
- [Qwen-Image](https://theaibench.ai/models/qwen-image/): Alibaba’s image models, strong at drawing readable text. 2512 is free for business; 2.1 is research-only.
- [HiDream-O1-Image](https://theaibench.ai/models/hidream-o1-image/): An 8B image model that works directly on pixels: generation, editing and personalised subjects. MIT licence.
- [Krea 2 Turbo](https://theaibench.ai/models/krea-2/): A 12B image model distilled to 8 steps. Free for business only under $1M yearly revenue.
- [Stable Diffusion](https://theaibench.ai/models/stable-diffusion/): Stability AI’s image models. SDXL is old but small and still the most downloaded; 3.5 Medium is newer.

## Speech models
- [Kokoro](https://theaibench.ai/models/kokoro-82m/): A tiny 82M voice model with 54 preset voices in 8 languages. Runs on almost anything, even without a graphics card.
- [Qwen3-TTS](https://theaibench.ai/models/qwen3-tts/): Alibaba’s voice models: clone a voice from 3 seconds, preset voices, or design a voice from a description. 10 languages.
- [Chatterbox](https://theaibench.ai/models/chatterbox/): Resemble AI’s voice-cloning models: Multilingual V3 covers 23 languages; Turbo is English-only and quicker.
- [VoxCPM2](https://theaibench.ai/models/voxcpm2/): A 2B voice model for 30 languages with high-quality 48 kHz audio, voice design and cloning.
- [OmniVoice](https://theaibench.ai/models/omnivoice/): A 0.6B voice model covering 600+ languages with cloning. The weights are non-commercial.
- [MOSS-TTS-Nano](https://theaibench.ai/models/moss-tts-nano/): A 100M voice model for 20 languages that its authors say runs on a 4-core CPU.
- [Supertonic 3](https://theaibench.ai/models/supertonic-3/): A small on-device voice model for 31 languages that runs without a graphics card.
- [Whisper large-v3-turbo](https://theaibench.ai/models/whisper/): OpenAI’s speech-to-text model for 99 languages, fast through faster-whisper. The default choice for transcription.
- [Parakeet TDT 0.6B v3](https://theaibench.ai/models/parakeet-tdt/): NVIDIA’s fast transcription model for 25 European languages, with word timings.
- [Qwen3-ASR](https://theaibench.ai/models/qwen3-asr/): Alibaba’s speech recognition for 30 languages and 22 Chinese dialects, in 1.7B and 0.6B sizes.
- [Nemotron 3.5 ASR Streaming](https://theaibench.ai/models/nemotron-asr/): NVIDIA’s low-delay streaming speech recognition for voice agents, 40 language-locales.
- [Voxtral Mini 4B Realtime](https://theaibench.ai/models/voxtral-realtime/): Mistral’s live transcription model for 13 languages with about half a second of delay.
- [Cohere Transcribe](https://theaibench.ai/models/cohere-transcribe/): Cohere’s 2B speech recognition for 14 languages, splitting long audio automatically.
- [VibeVoice-ASR](https://theaibench.ai/models/vibevoice-asr/): Microsoft’s 8.7B model that transcribes up to an hour in one pass, with who-said-what and timings.

## Embedding models
- [Embedding models](https://theaibench.ai/models/embeddings/): Small models that turn text into numbers for search and “chat with your documents”. Most run on any computer.