- Family
- DeepSeek R1
- Context
- 128K
- Parameters
- 671B
- Architecture
- Moe
- Weights
- Downloadable
- License
- MIT
Open Models Research
Compare open-weight models without treating every checkpoint as a separate product.
This view is generated from iilib's shared model graph. It focuses on representative, source-backed models and separates downloadable weights, licensing, commercial-use terms, architecture, self-hostability and research transparency.
“Open-weight” does not automatically mean “open source.” Custom community and vendor licenses remain visible instead of being collapsed into a single open/closed label.
Openness signals
Weights are only one part of the picture
4 models explicitly expose source code or recipes; 2 expose public training data; 2 expose training checkpoints. Missing transparency fields are not inferred from marketing language.
Method
Family, developer and hosted API availability are derived from verified graph relationships. Use-case labels are editorial research taxonomy, not benchmark rankings.
Read the research methodology →Research table
Filter the current representative set
Filters run only in your browser and do not create indexable URL combinations. The default table remains the canonical research view.
- Family
- DeepSeek V3
- Context
- 128K
- Parameters
- 671B
- Architecture
- Moe
- Weights
- Downloadable
- License
- MIT
- Family
- Gemma 4
- Context
- 256K
- Parameters
- 11.95B
- Architecture
- Dense
- Weights
- Downloadable
- License
- Apache-2.0
- Family
- Gemma 4
- Context
- 256K
- Parameters
- 25.2B
- Architecture
- Moe
- Weights
- Downloadable
- License
- Apache-2.0
- Family
- GLM-5.3
- Context
- 1.05M
- Parameters
- 744B
- Architecture
- Moe
- Weights
- Downloadable
- License
- GLM-5.3 License
- Family
- GLM-5.3
- Context
- 1.05M
- Parameters
- 320B
- Architecture
- Hybrid
- Weights
- Downloadable
- License
- MIT
- Family
- gpt-oss
- Context
- 131K
- Parameters
- 116.83B
- Architecture
- Moe
- Weights
- Downloadable
- License
- Apache-2.0
- Family
- Kimi K3
- Context
- 1.05M
- Parameters
- 2,800B
- Architecture
- Moe
- Weights
- Downloadable
- License
- Kimi K3 License
- Family
- Llama 4
- Context
- 1M
- Parameters
- 400B
- Architecture
- Moe
- Weights
- Restricted
- License
- Llama 4 Community License
- Family
- Llama 4
- Context
- 10M
- Parameters
- 109B
- Architecture
- Moe
- Weights
- Restricted
- License
- Llama 4 Community License
- Family
- Mistral Medium
- Context
- 256K
- Parameters
- 128B
- Architecture
- Dense
- Weights
- Downloadable
- License
- Modified MIT
- Family
- Nemotron 3.5
- Context
- 1M
- Parameters
- 30B
- Architecture
- Hybrid
- Weights
- Downloadable
- License
- OpenMDW-1.1
- Family
- Olmo 3
- Context
- 66K
- Parameters
- 7B
- Architecture
- Dense
- Weights
- Downloadable
- License
- Apache-2.0
- Family
- Olmo 3
- Context
- 66K
- Parameters
- 32B
- Architecture
- Dense
- Weights
- Downloadable
- License
- Apache-2.0
- Family
- Qwen3
- Context
- 128K
- Parameters
- 235B
- Architecture
- Moe
- Weights
- Downloadable
- License
- Apache-2.0
- Family
- Qwen3
- Context
- 128K
- Parameters
- 30B
- Architecture
- Moe
- Weights
- Downloadable
- License
- Apache-2.0
- Family
- Qwen3
- Context
- 32K
- Parameters
- 4B
- Architecture
- Dense
- Weights
- Downloadable
- License
- Apache-2.0
- Family
- Qwen3 Embedding
- Context
- 32K
- Parameters
- 0.6B
- Architecture
- Dense
- Weights
- Downloadable
- License
- Apache-2.0
- Family
- Qwen3 Reranker
- Context
- 32K
- Parameters
- 0.6B
- Architecture
- Dense
- Weights
- Downloadable
- License
- Apache-2.0
What is intentionally missing?
The table does not mirror every parameter size, quantization, GGUF/AWQ/GPTQ conversion, community fine-tune, hosted alias or intermediate training checkpoint. A new Model record must add a meaningful research dimension.
No universal “best model” score
Benchmark leaderboards are excluded from this first research layer because evaluation versions, harnesses, prompts, contamination risk and provider settings differ. Structured facts remain separate from future benchmark evidence.