BenchLM recommendation
Best Ollama Models (July 2026): Ranked + Pull Commands
The best Ollama model for most machines is Qwen3.6-27B (BenchLM 65): a ~16GB Q4_K_M pull that fits one 24GB GPU and posts a 77.2 SWE-bench Verified, making it the practical coding pick too. The highest-scoring model you can serve through Ollama is Kimi K2.6 (74), but its ~600GB Q4 weights need a multi-GPU server. On an 8GB card, run Mellum2-12B-A2.5B-Thinking (59).
Ollama is the easiest way to run a large language model on your own hardware: one ollama pull downloads a quantized GGUF build of a model, and ollama run serves it locally with an OpenAI-compatible API — on macOS, Windows, or Linux, with no CUDA setup or Python environment. By default the library ships models at Q4_K_M quantization, a 4-bit format that cuts weights to roughly 0.6GB per billion parameters with a modest quality cost.
That means only open-weight modelsqualify for this page — models whose weights are publicly downloadable. Proprietary APIs like GPT-5.4 or Claude can never run in Ollama. Every score below is the model's BenchLM display score, the same provisional weighted average used across our leaderboard, so you can compare these picks directly against our open-source ranking and the hardware-tiered best local LLMs guide. Parameter counts come from BenchLM's self-host catalog or the official model names; Q4_K_M sizes are estimates from parameter count (~0.6GB per billion), not measured downloads.
The best Ollama models, ranked
Ten open-weight models worth pulling in July 2026, sorted strictly by BenchLM score. The top two are frontier MoE models that technically run through Ollama but realistically need datacenter hardware; everything from Qwen3.6-27B down fits a single consumer GPU or a Mac with enough unified memory.
| Model | BenchLM score | Params | Est. Q4_K_M size | Min hardware (Q4) | Pull command |
|---|---|---|---|---|---|
| Kimi K2.6Moonshot AI | 74 | 1000B MoE (32B active) | ~600GB | 8× NVIDIA H100 (80GB) | ollama pull kimi-k2 |
| GLM-5.1Z.AI | 68 | 744B MoE (40B active) | ~446GB | 8× NVIDIA H100 (80GB) | ollama pull glm |
| Qwen3.6-27BAlibaba | 65 | 27B dense | ~16GB | NVIDIA RTX 4090 (24GB) | ollama pull qwen3.6 |
| Kimi K2.5Moonshot AI | 63 | 120B dense | ~72GB | 4× NVIDIA A100 (80GB) | ollama pull kimi-k2 |
| Gemma 4 31BGoogle | 61 | 31B dense | ~19GB | NVIDIA RTX 4090 (24GB) | ollama pull gemma4 |
| Mellum2-12B-A2.5B-ThinkingJetBrains | 59 | 12B MoE (2.5B active) | ~7GB | 8GB GPU or 16GB Mac | ollama pull mellum |
| Qwen3.6-35B-A3BAlibaba | 58 | 35B MoE (3B active) | ~21GB | 24GB GPU or 32GB Mac | ollama pull qwen3.6 |
| Gemma 4 12BGoogle | 52 | 12B dense | ~7GB | 8–12GB GPU or 16GB Mac | ollama pull gemma4 |
| Mistral Small 4Mistral | 49 | 119B MoE (22B active) | ~71GB | 2× NVIDIA RTX 5090 (32GB GDDR7) | ollama pull mistral-small |
| DeepSeek-R1DeepSeek | 33 | 671B MoE (37B active) | ~403GB | 8× NVIDIA H100 (80GB) | ollama pull deepseek-r1:671b |
Pull commands use each model's common Ollama library family name. Newer releases often add versioned or size-suffixed tags (like deepseek-r1:671b) — check ollama.com/library for the exact tag and available sizes before pulling. Q4_K_M sizes are estimates from parameter count.
Best Ollama model for coding
On a single 24GB GPU, Qwen3.6-27B is the best coding model you can pull: 77.2 on SWE-bench Verified and 83.9 on LiveCodeBench from a dense 27B whose Q4_K_M build is only ~16GB, with a 262K context for repository-scale work. If you have server-class hardware, Kimi K2.6is the outright coding leader among open weights — 80.2 SWE-bench Verified and 89.6 LiveCodeBench — but its 1T-parameter MoE demands roughly 600GB of memory at Q4. On an 8GB card, JetBrains' Mellum2-12B-A2.5B-Thinking (69.9 LiveCodeBench) is a purpose-built coding MoE with just 2.5B active parameters, so it stays fast even partially offloaded to CPU. Pair any of them with Ollama's local API and your editor's OpenAI-compatible endpoint setting.
Best small Ollama models (8–16GB)
With 8–16GB of VRAM or unified memory you are shopping in the 8–14B class, where Q4_K_M builds land around 5–8GB. Mellum2-12B-A2.5B-Thinking (59) is the standout: a ~7GB pull that scores within six points of models needing a 24GB card. Google's Gemma 4 12B(52) is the best general-purpose dense option at the same ~7GB size, and Zyphra's ZAYA1-8B (55) is the strongest pure 8B, though it is not in the Ollama library — import its GGUF from Hugging Face instead. For reasoning on small cards, the distilled deepseek-r1 tags ship sizes from 7B to 70B.
How to choose an Ollama model
- Start from your memory, not the leaderboard. At Q4_K_M, budget ~0.6GB per billion parameters for weights, then leave 20–30% headroom for context (KV cache). A 24GB card comfortably runs a ~16GB model; it will not run a ~21GB one at useful context lengths.
- Prefer MoE models when you must offload to CPU. Sparse models like Mellum2 (2.5B active) or Qwen3.6-35B-A3B (3B active) only compute a fraction of their weights per token, so they stay usable on Macs and mixed CPU/GPU setups where dense models crawl.
- Pick by task, not just overall score. For coding, weight SWE-bench Verified and LiveCodeBench; Qwen3.6-27B beats several models that outscore it overall.
- Check the license.Apache-2.0 (Qwen, Gemma's terms differ slightly) and MIT (GLM-5.1) allow commercial use; some models ship modified or research licenses.
- Verify the exact tag before pulling. Library names change between releases — confirm on ollama.com/library, and prefer explicit size tags over defaults.
- Outgrowing one machine? Price a dedicated GPU setup with our self-host cost calculator before renting API capacity.
Ollama model FAQ
What is the best Ollama model right now?
Kimi K2.6 is the highest-scoring open-weight model you can serve through Ollama, at 74 on BenchLM's July 2026 leaderboard — but its 1T-parameter MoE needs roughly 600GB of memory at Q4, firmly server territory. For a machine you actually own, Qwen3.6-27B (65) is the best answer: its ~16GB Q4_K_M build fits a single 24GB GPU or a 32GB Mac and leads its size class on coding benchmarks.
What is the best Ollama model for coding?
Qwen3.6-27B is the best coding model for typical Ollama hardware, scoring 77.2 on SWE-bench Verified and 83.9 on LiveCodeBench from a single 24GB GPU. Kimi K2.6 leads all open weights (80.2 SWE-bench Verified, 89.6 LiveCodeBench) if you have a multi-GPU server. On an 8GB card, JetBrains' Mellum2-12B (69.9 LiveCodeBench) is the strongest dedicated coding pull.
How much RAM or VRAM do I need for Ollama?
Budget roughly 0.6GB per billion parameters at Q4_K_M, plus 20–30% headroom for context. In practice: 8GB runs 7–12B models (Gemma 4 12B, Mellum2), 16GB handles the 14–20B class, and 24GB fits 27–31B picks like Qwen3.6-27B and Gemma 4 31B. Macs count unified memory the same way. Beyond ~32GB you are into multi-GPU rigs or high-memory Mac Studios — see our local LLM guide by VRAM tier.
Are Ollama models free?
Yes. Ollama itself is free, open-source software, and every model in its library is open-weight, so downloading and running them costs nothing beyond your hardware and electricity — no per-token fees. Licenses still vary: Qwen3.6-27B and Mistral Small 4 are Apache-2.0 and GLM-5.1 is MIT (all commercial-friendly), while models like Kimi K2.6 use modified licenses. Check the license before shipping a commercial product on top of one.
Local LLM updates
Get notified when a new open-weight release changes the Ollama rankings or drops the hardware bar.
Free. No spam. Unsubscribe anytime.