Top 17 open models you can run yourself

The highest-ranked model with downloadable weights on this board

Best for Teams with a large cluster who want the best-ranked open modelYou give up Needs over 1 TB of memory; API costs like a closed model

Top open-weights model on the Artificial Analysis index, at cents per task; closed leaders still rank well above it.

Best for Coding and agent work for cents per task; reads text, images, video and audioYou give up Behind closed leaders on coding and agents; Xiaomi's platform stores data in China

Open-weights general model with fair API prices

Best for Agent coding and security work with open weightsYou give up MiMo-V2.6-Pro costs far less per task for similar marks

Very cheap and openly licensed; good for high-volume everyday chat

Best for Bulk API jobs with image input, cheapest in off-peak hoursYou give up Peak hours cost double; the board still rates the earlier version

Very cheap MIT model, 13th on arena.ai Math; for coding and agents MiMo-V2.6-Pro ranks higher for less per task.

Best for Maths-heavy or bulk work at one of the lowest prices hereYou give up Mid-table for coding and weak at creative writing

Larger DeepSeek model; the cheaper V4.1 Flash ranks higher on this board

Best for Agent coding through DeepSeek's API, at half price off-peakYou give up Far too large to self-host for most teams; V4.1 Flash ranks higher for less

Apache-2.0 295B model (21B active); Tencent suggests 8 large-memory GPUs to serve it

Best for Self-hosting under a permissive licence; cheaper to serve than Tencent's Hy4You give up Weak at web app coding and maths; 256K context, shorter than most here

Google's biggest Gemma; Google says it targets consumer GPUs and workstations. Apache-2.0

Best for One 32 GB graphics card or a workstation, with image inputYou give up Ranks low in chat votes, well below the larger open models

One of the cheapest APIs here with MIT weights; 309B parameters (15B active), so for servers

Best for Very cheap API work with text, image, video and audio inputYou give up MiMo-V2.6-Pro scores higher (87 against 79) for a few cents more per task

Low API price; 428B parameters, so for servers. Prompts over 512K tokens cost double

Best for Long coding tasks with image and video input at a low priceYou give up Custom community licence; GLM-5.3-Flash ranks higher on web app coding for less

Apache-2.0 975B model (41B active); a job for teams with a GPU cluster

Best for Teams that want a large Apache-2.0 base model to fine-tuneYou give up Low in chat votes for its size; third-party API about $4.68 per 1M output

Apache-2.0 27B dense model, far smaller to host than the open leaders above it

Best for One high-end GPU through Ollama or LM Studio; a popular base to fine-tuneYou give up Ranks well below the open leaders; thinking is on by default

Only 3.8B of 25.2B parameters active; Google says it targets consumer GPUs. Apache-2.0

Best for Fast local use on a laptop graphics card with about 17 GB of memoryYou give up Ranks low in chat votes, well below the larger open models

European 128B dense model; licence excludes very large companies, and Mistral's example uses 8 GPUs

Best for Smaller firms that want a European model with image inputYou give up Low on both boards; third-party API about $7.50 per 1M output

NVIDIA's 550B model, fine for commercial use; NVIDIA lists 8 B200 or 16 H100 GPUs as the minimum

Best for Teams that want open training recipes and a 1M token contextYou give up Ranks below the Chinese open leaders on quality

Meta's Apache-2.0 model; its 4-bit version targets 24GB or 32GB consumer GPUs

Best for Agent tasks on one 24 GB consumer graphics card, under Apache-2.0You give up Low in chat votes (108th of distinct models); reads images but not audio

121B model that reads images; few votes yet, so its rank may move. Needs a data-center GPU

Best for Commercial image-reading tasks on one large data-center GPUYou give up Low in chat votes (118th of distinct models); no review yet