GLM-5.2 leads July's open-weight LLMs as Inkling tops US models

Kimi K3 and Qwen3.8 remain outside the ranking because their downloadable weights were not available by July 20th.

By · Published

Why it matters

Open-weight adoption depends on downloadable checkpoints and licenses, not launch claims. GLM-5.2 leads today, while Inkling gives US companies their strongest domestic option.

GLM-5.2 leads July's open-weight LLMs as Inkling tops US models — Kimi K3 and Qwen3.8 remain outside the ranking because their downloadable weights were not available by July 20th.

Z.ai's GLM-5.2 holds the lead among downloadable open-weight language models as of July 20th, scoring 51 on the Artificial Analysis Intelligence Index and opening a seven-point gap over its nearest tested rivals. The 753-billion-parameter model is available under the MIT license, activates roughly 40 billion parameters per token and supports a 1-million-token context window. (huggingface.co)

The ranking also shows how thoroughly Chinese labs control the open-weight market. Models from Z.ai, MiniMax, DeepSeek, Moonshot AI, Xiaomi and Tencent occupy eight of the 10 places. Nex-N2-Pro, another entry, was post-trained on Alibaba's Qwen3.5 architecture. Thinking Machines Lab's Inkling is the sole US-built model in the group and the highest-ranked American release. (huggingface.co)

Inkling gives Mira Murati's Thinking Machines Lab a credible first entry in a market where US frontier labs have largely kept their strongest model weights private. Released on July 15th under Apache 2.0, Inkling has 975 billion total parameters, activates 41 billion and accepts text, image and audio inputs across a 1-million-token context window. Thinking Machines says it trained the model from scratch on 45 trillion multimodal tokens. (thinkingmachines.ai)

The top 10 open-weight LLMs

The list below uses each model's highest tested reasoning configuration on the Artificial Analysis Intelligence Index, rounded to the nearest whole point. It counts each underlying model once, rather than treating different API providers or free endpoints as separate entries. The index combines agentic work, coding, scientific reasoning, knowledge reliability and instruction-following evaluations. Rankings can move as evaluators rerun models or revise their test suites. (modelgrep.com)

  1. GLM-5.2 - 51. Z.ai's MIT-licensed flagship is the clear leader. Its strongest independent results came in long-horizon agent work, though it also used more output tokens per task than MiniMax-M3, Kimi K2.6 and DeepSeek V4 Pro. (huggingface.co)

  2. MiniMax-M3 - 44. MiniMax-M3 combines image and video input with a 1-million-token context window. Its 428-billion-parameter architecture activates 23 billion parameters per token. The weights carry MiniMax's community license, which requires a separate agreement for some commercial use. (artificialanalysis.ai)

  3. DeepSeek V4 Pro - 44. DeepSeek's larger V4 model uses 1.6 trillion total parameters and activates 49 billion. Its MIT license and comparatively low API prices make it one of the strongest options for companies balancing capability, deployment control and inference cost. (artificialanalysis.ai)

  4. Kimi K2.6 - 44. Moonshot AI's K2.6 is a 1-trillion-parameter multimodal model with 32 billion active parameters and a 256,000-token context window. Its benchmark performance remains competitive, though independent testing found it slower and more verbose than several similarly ranked models. (artificialanalysis.ai)

  5. MiMo-V2.5-Pro - 42. Xiaomi's MiMo-V2.5-Pro has 1.02 trillion total parameters, 42 billion active parameters and a 1-million-token context window. Xiaomi released the weights under MIT, making it one of the least commercially restrictive trillion-parameter models in the group. (huggingface.co)

  6. Kimi K2.7 Code - 42. Moonshot's coding specialist shares K2.6's 1-trillion-parameter architecture but is tuned for software engineering and long-running coding agents. Moonshot says the release cut reasoning-token use by about 30% from K2.6, while independent testing places it near MiMo-V2.5-Pro on overall intelligence. (kimi.com)

  7. Tencent Hy3 - 41. Hy3 is a 295-billion-parameter model with about 21 billion active parameters, a 256,000-token context window and an Apache 2.0 license. Its smaller active footprint makes it easier to serve than the trillion-parameter entries, although Tencent still recommends an eight-GPU system for deployment. (huggingface.co)

  8. Nex-N2-Pro - 41. Nex-AGI's agent model is built on Qwen3.5-397B-A17B and focuses on coding, research and terminal-based work. The architecture gives Nex-AGI a faster route into the top group by post-training an existing Qwen base rather than pretraining a frontier model from zero. (huggingface.co)

  9. Inkling - 41. Thinking Machines' first open-weight model ranks alongside Hy3 and Nex-N2-Pro while adding native image and audio understanding. Its position matters because it gives US companies an Apache-licensed alternative to a field dominated by Chinese checkpoints. (artificialanalysis.ai)

  10. DeepSeek V4 Flash - 40. V4 Flash activates only 13 billion of its 284 billion parameters. Artificial Analysis measured roughly 110 output tokens per second and first-party pricing of $0.14 per million input tokens and $0.28 per million output tokens, making Flash the cost-focused model in DeepSeek's two-tier V4 lineup. (artificialanalysis.ai)

Why Kimi K3 and Qwen3.8 are missing

Two models attracting the most attention in July do not qualify for this ranking yet. Moonshot released Kimi K3 through its apps and API on July 16th, claiming 2.8 trillion parameters, native vision and a 1-million-token context window. Moonshot scheduled the full weight release for July 27th. On July 20th, developers could rent access to K3 but could not download the promised checkpoint. (kimi.com)

Alibaba's Qwen3.8-Max-Preview has the same problem. Alibaba says the preview has 2.4 trillion parameters and plans to release open-weight versions, but the launch provided no downloadable Max checkpoint, model card, activated-parameter count or independent benchmark result. Until those assets land, Qwen3.8 is a proprietary preview carrying a future open-weight commitment. (scmp.com)

That distinction is material for engineering teams. API availability lets developers test a model. Public weights and a usable license determine whether they can inspect it, fine-tune it, run it behind their own security boundary or move between infrastructure providers. The July ranking therefore favors models that developers can deploy today over models whose makers have promised to open them later.

Reader comments

Conversation for this story loads after sign-in.