There is no single best open model — ten open-weight picks, each best at one job
Asked:
“Best opensource Ai models”
Ten recommended models from 8 independent sources, as of 9 November 2026. Most are open-weight, not fully open-source: the weights are downloadable, but training data, code, and usage freedoms vary. Apache 2.0 and MIT-family terms are permissive; Mistral Large 4 requires a separate commercial license for Model-as-a-Service businesses — verify each license before deploying. Benchmarks are vendor-reported or drawn from comparison guides and are not directly comparable, so this is a map of “best for” choices, not a 1-to-10 ranking.
Size vs context: where each model fits
Total parameters (billions, log scale) against context window (tokens, log scale). Colour is the use case; a hollow ring marks a sparse mixture-of-experts model (few active parameters, cheaper to run than its total size suggests). Mistral Large 4, GLM-Image and Qwen 3.7 lack a stated context window in the sources and sit on the bottom shelf.
Mistral Large 4 (Le Chonk) is the frontier agentic coding and cybersecurity pick — 48% Coding Agent Index, 51% Cyber Index — but Model-as-a-Service businesses need a separate commercial license. mistral.ai
For complex coding agents, DeepSeek V4-Pro (80.6% SWE-Bench Verified, MIT) and Kimi K2.6 (80.2%, modified MIT) are near-tied at the top. qubrid.com
On a high-end workstation, Qwen3.8-27B is the practical multimodal choice: 27.78B parameters, a 262,144-token context, Apache 2.0. kingy.ai
gpt-oss-20b runs reasoning and tool use locally on a small budget — 20.9B total but only 3.6B active parameters, Apache 2.0 plus OpenAI's usage policy. cdn.openai.com
Kimi K3 is the outlier for huge-context work: a 1,048,576-token window on a 2.8T-total multimodal model for long-horizon agentic research. tech-insider.org
Model
Developer
Category
Best for
Parameters
Context
License
Released
Source
Source: 10-model open-weight shortlist compiled from 8 independent hosts, retrieved 9 November 2026; one row per recommendation. Numbers are vendor-reported or from comparison guides under differing harnesses and are not cross-comparable. Blank fields mean the source did not clearly state the value, not that the model lacks the feature. Full benchmark evidence per model appears in chart tooltips; long license texts shortened in the table.