Model verification
Source, license, context, and availability notes for fast-moving model pages. Check the current model card before adoption.
License/status: MIT license for DeepSeek-V4 preview weights; verify the exact checkpoint before redistribution or commercial deployment.
Openness: Open weights
Parameters: DeepSeek-V4-Pro: 1.6T total parameters, 49B active parameters per token, MoE architecture.
Context: Reported 1M-token context window for the V4 preview series.
Verified through: June 2026
License/status: Apache 2.0 for Qwen3-235B-A22B public model-card release.
Openness: Open weights
Parameters: 235B total parameters with 22B active parameters per token.
Context: Base Qwen3 model-card context support varies by release; newer Instruct releases list 262K native context with long-context extension paths.
Verified through: June 2026
License/status: Modified MIT for Kimi K2 open-weight releases where the model card identifies that license; hosted Kimi app/API terms may differ.
Openness: Open weights for released K2 checkpoints; hosted service is commercial/proprietary.
Parameters: Kimi K2 family uses MoE architecture; verify exact K2.6 checkpoint parameter and active-parameter counts from Moonshot/Hugging Face before publishing benchmark claims.
Context: Kimi K2 0905 doubled context from 128K to 256K; verify K2.6 context on the current Moonshot model card/API docs.
Verified through: June 2026
License/status: MIT for GLM open-weight releases listed by Z.ai/Zhipu where the exact model card uses that license; verify the specific GLM-5.1 checkpoint.
Openness: Open weights where released by Z.ai; hosted/API terms may differ.
Parameters: GLM-5.1 is tracked as a flagship Z.ai model; verify exact parameter count and variant on the Z.ai model card before publishing specs.
Context: Context limits vary by GLM release and provider endpoint; verify current Z.ai docs.
Verified through: June 2026
License/status: Treat MiMo-V2.5-Pro as provider/API or source-status-restricted unless Xiaomi publishes an exact open-weight checkpoint and license for this Pro release.
Openness: Not confirmed as open weights for the Pro checkpoint in this directory; MiMo Flash/open variants may have different licenses.
Parameters: MiMo Pro-class releases are reported around 1T+ total parameters with 42B active parameters; verify exact V2.5-Pro numbers from Xiaomi before production claims.
Context: Pro-class MiMo releases have been reported with up to 1M context; verify exact checkpoint/provider limits.
Verified through: June 2026
License/status: Apache 2.0 for Gemma 4 where Google model cards identify that license; Gemma 3 and earlier used Gemma Terms of Use/source-available terms.
Openness: Open models/open weights; verify exact Gemma variant and license.
Parameters: Gemma 4 family includes efficient small variants plus larger dense/MoE variants; verify exact model size on the Google/Kaggle/Hugging Face card.
Context: Context limits vary by Gemma generation and size; verify the exact Gemma 4 checkpoint.
Verified through: June 2026
License/status: License/source availability must be verified against the official MiniMax M2.7 model page before commercial or redistribution claims.
Openness: Model information public; exact open-weight/license status should be verified per checkpoint.
Parameters: MiniMax-M2 series paper describes flagship M2 as 229.9B total parameters with 9.8B active parameters; M2.7 is described as a later checkpoint in the series.
Context: Verify context window and provider limits on the current MiniMax model page/API docs.
Verified through: June 2026
License/status: Apache 2.0 for weights, inference implementations, tool environments, and tokenizer per OpenAI model-card release.
Openness: Open weights
Parameters: 117B total parameters with 5.1B active parameters per token.
Context: 128K/131K-class context depending on runtime and source wording; verify runtime-specific limits.
Verified through: June 2026
License/status: Apache 2.0 for weights, inference implementations, tool environments, and tokenizer per OpenAI model-card release.
Openness: Open weights
Parameters: 21B-class model with 3.6B active parameters per token; designed for lower-memory local deployment.
Context: 128K/131K-class context depending on runtime and source wording; verify runtime-specific limits.
Verified through: June 2026
Qwen family pages should verify exact checkpoint license and context limits because Base, Instruct, Thinking, Coder, VL, Embedding, and provider variants differ.
Verified through: June 2026
DeepSeek family pages should distinguish V4, R1, coder, and distilled releases. Exact license, context, and hardware requirements must be read from the current checkpoint card.
Verified through: June 2026
Kimi family pages should distinguish open-weight K2 checkpoints from the hosted Kimi product/API terms.
Verified through: June 2026
GLM family pages should verify exact Z.ai checkpoint names because GLM release status, API terms, and context windows differ by variant.
Verified through: June 2026
Gemma family pages should separate Gemma 4 Apache-licensed checkpoints from earlier Gemma releases that used Gemma-specific terms.
Verified through: June 2026
MiMo family pages should not assume Pro checkpoints are open-weight unless Xiaomi publishes the exact weights and license. Flash/open variants may have different terms.
Verified through: June 2026
MiniMax family pages should verify exact M2/M2.7 checkpoint license, context, and availability from the official MiniMax model page before open-weight claims.
Verified through: June 2026
gpt-oss pages should cite the OpenAI model card paper and distinguish 120B/20B hardware requirements, active parameters, and runtime context limits.
Verified through: June 2026