Seven Seats in Fifty: only seven of the 50 highest-rated models on Arena have weights anyone can download, and all seven come from Chinese labs. The best of them, Kimi K3.1 Pro, trails the top closed model by 19 points on Artificial Analysis' index, at a fraction of the price.
Open-source LLM leaderboard vs. closed models: each dot is one of the 50 highest-rated models on Arena's text leaderboard, stacked by score. Open-weight models are highlighted, proprietary ones gray. The first open seat is #20.
Arena text leaderboard, last updated Nov 11, 2026, ranked by score; ties keep Arena's order. Scores come from crowd votes between anonymous pairs of answers, with confidence intervals of ±3 to ±10, so models a few points apart are statistically tied. "Open weights" means Arena lists a non-proprietary license; Kimi models use Moonshot's own licenses, the rest MIT. All seven open models are from Chinese labs (Moonshot, Z.ai, Xiaomi, DeepSeek).
Open-source vs. closed LLM comparison on the Artificial Analysis Intelligence Index (v5), Nov 13, 2026. The first four rows are the top closed models; the rest are the top six open-weight models.
Each model at its best-scoring setting: Claude Fable 5.2, Claude Opus 5.5 and GPT-6 Astra at max, Gemini 4 Carbon at high, DeepSeek V4.1 Flash 1017 and GLM-5.3 at Max, Mistral Large 4 at High. Closed scores are from Artificial Analysis' leaderboard and open scores from its open-weights page, both read Nov 13, 2026. Ling 3.1 Flash 1019 also scores 30. Version 5 of the index, launched Oct 27, scores harder tasks than earlier versions, so these numbers can't be compared with spring 2026 scores. At launch, AA put the open-weights gap at 12 points (Opus 5.5 at 58 vs. Kimi K3.1 Pro at 46).
Cheapest LLM API pricing vs. frontier models: list price per 1 million output tokens, US$, on each lab's own API, Nov 13, 2026. The bottom three are open-weight models.
Standard list prices for uncached output. DeepSeek's are peak-hour rates; off-peak is half. "DeepSeek Flash" is the "deepseek-flash" API, which serves DeepSeek-V4.1-Flash-1017; "deepseek-v4-pro" serves V4-Pro-0813. Input prices follow the same order (DeepSeek V4.1 Flash $0.30, GLM-5.3 $1.40, Claude Opus 5.5 $4, GPT-6 Astra $10). Open weights can also be self-hosted, at whatever your hardware costs. The 1/33 ratio is our calculation ($1.20 ÷ $40). Moonshot's price table for Kimi K3.1 Pro did not load, so it is left out.
DeepSeek V4 Pro benchmark scores, non-thinking vs. Max reasoning mode, as reported by DeepSeek, %
Vendor-reported scores from DeepSeek's own table (Pass@1, EM or resolved rate). Max mode spends far more reasoning tokens per answer. Most open-model rankings, including the article this graphic checks, quote each model's best mode without saying so. The axis starts at 50.
Chinese vs. U.S. open-source LLM downloads: cumulative Hugging Face downloads by model origin, through March 2026, millions
About 1,500 mainline open language models; a download is any HTTP request for a model file. China overtook the U.S. in late July 2025 and grew 11.9× in a year (97M to 1.15B) vs. 4.1× for the U.S. The ATOM Report puts all tracked downloads at 2.04 billion.
Open-weight model sizes: active parameters as a share of total parameters in mixture-of-experts models, %
Our calculation from each vendor's card: Kimi K3.1 Pro 104B of 2.8T, Inkling 41B of 975B, GLM-5.3 40B of 744B, MiniMax M3 ~23B of ~428B, Kimi K2.7 Code 32B of 1T, DeepSeek V4 Pro 49B of 1.6T. All of the weights must still fit in memory. Gemma 4 31B is dense (30.7B), so every parameter is active.