Best Local LLM (November 2026) — Open-Weights Models by Size
As of Nov 22, 2026, the best open-weights LLM you can run yourself is Kimi K3 (2800B parameters), #1 among open-weights models on isitsota's General LLM board. At 32B or fewer, the best is Gemma 4 31B (31B).
Last updated Nov 22, 2026 · Full open-weights General LLM board
Best pick by size
- 15B to 32B: Gemma 4 31B (31B, #12 among open-weights models)
- Over 80B: Kimi K3 (2800B, #1 among open-weights models; 15 ranked in this size)
Ranked open-weights models with a published size
| Rank | Model | Parameters | License | Coding rank |
|---|---|---|---|---|
| 1 | Kimi K3 | 2800B | Proprietary | #5 |
| 1 | GLM 5.3 | 744B | MIT | #1 |
| 1 | Mimo V2.6 Pro | 1020B | MIT | #4 |
| 2 | GLM 5.3 Flash | 320B | MIT | #1 |
| 4 | DeepSeek V4.1 Flash | 552B | MIT | #2 |
| 5 | Mimo V2.5 Pro | 1020B | MIT | — |
| 6 | GLM 5.1 | 754B | MIT | — |
| 7 | DeepSeek V4 Pro | 1600B | MIT | — |
| 8 | Kimi K2 6 | 1040B | Modified MIT | — |
| 9 | GLM 5 | 744B | MIT | — |
| 12 | Gemma 4 31B | 31B | Apache 2.0 | — |
| 13 | Kimi K2 5 | 1040B | Modified MIT | — |
| 15 | Inkling | 975B | Apache 2.0 | #11 |
| 31 | GLM 4.6 | 357B | MIT | — |
| 44 | GLM 4.5 | 355B | MIT | — |
| 51 | DeepSeek R1 | 671B | MIT | — |
Rank is the position among open-weights models on the General LLM board; coding rank is the same on the Coding board. Parameters are total parameters as published; mixture-of-experts models use fewer per token, and memory needs depend on quantization. 7 more ranked open-weights models have no published parameter count and are not listed.
What is the best local LLM?
Kimi K3 (2800B parameters) ranks highest among open-weights models on isitsota's General LLM board as of Nov 22, 2026 (#1).
What is the best local LLM under 32B parameters?
Gemma 4 31B (31B), #12 among open-weights models on the General LLM board as of Nov 22, 2026.
What is the best local LLM for coding?
GLM 5.3 (744B) ranks #1 among open-weights models on the Coding board as of Nov 22, 2026.
How is this list ranked?
By the same benchmark-weighted General LLM board used across isitsota, limited to open-weights models with a published parameter count. Counts are total parameters; mixture-of-experts models use fewer per token, and memory needs depend on quantization.