Skip to main content
BenchLM
Data
Recommendation
Ranking data as of

Best Ollama Models (November 2026): Ranked + Pull Commands

The best Ollama model for a single 24GB GPU is Qwen3.8-27B (BenchLM score 54.06): an 18GB download that runs on a 24GB GPU or a 32GB Mac. The highest-scoring model you can pull and run locally is Ornith-1.5-397B (61.41), but its 242GB download needs 4× 80GB GPUs or a 512GB Mac. On an 8GB card, run Ornith-1.5-9B (32.5).

Ollama is the easiest way to run a large language model on your own hardware: one ollama pull downloads a quantized 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. Most library tags ship at Q4_K_M, a 4-bit format that cuts weights to roughly 0.6GB per billion parameters with a modest quality cost.

Only open-weight models qualify for this page, and only those with a local build in the Ollama library. Every score comes from the same public lane as the main leaderboard, so you can compare these picks directly against our open-source ranking and the hardware-tiered best local LLMs guide. Download sizes and context windows are Ollama's own figures for each tag, read from ollama.com/library on November 12, 2026.

The best Ollama models, ranked

37 open-weight models with a BenchLM score that you can pull from the Ollama library, sorted by score. Minimum hardware assumes the download plus about 20% headroom for context; Macs are sized at three quarters of unified memory, the share macOS gives the GPU.

Looking for the whole library? All 269 Ollama models are listed below with download sizes, the memory each one needs, and its score where we rank it.

ModelScoreDownloadContextMin hardwarePull command
Ornith-1.5-397BOrnith AI61.41242GB256K4× 80GB GPUs or 512GB Macollama pull ornith-1.5:397b
Qwen3.8-27BAlibaba54.0618GB256K24GB GPU or 32GB Macollama pull qwen3.8:27b
Qwen3.8-Flash-NextAlibaba54.06120GB256K2× 80GB GPUs or 192GB Macollama pull qwen3.8-flash-next:125b-a6b-q4_K_M
Qwen3.5-122B-A10BAlibaba52.281GB256K2× 80GB GPUs or 192GB Macollama pull qwen3.5:122b-a10b
Ternary Bonsai 2 27BPrism ML51.668.3GB256K12GB GPU or 16GB Macollama pull bonsai-2:27b
Nemotron 3.5 Super VL 120B A12BNVIDIA48.7880GB1M2× 80GB GPUs or 128GB Macollama pull nemotron-3.5-super:120b
Qwen3.6-27BAlibaba48.0518GB256K24GB GPU or 32GB Macollama pull qwen3.6:27b
Qwen3.5-27BAlibaba47.9417GB256K24GB GPU or 32GB Macollama pull qwen3.5:27b
Qwen3.5-35B-A3BAlibaba45.2522GB256K32GB GPU or 48GB Macollama pull qwen3.5:35b-a3b
Qwen3.6-35B-A3BAlibaba42.0123GB256K32GB GPU or 48GB Macollama pull qwen3.6:35b-a3b
Nemotron 3.5 Bolt 74B-A9BNVIDIA41.6125GB256K32GB GPU or 48GB Macollama pull nemotron-3.5-bolt:74b-a9b
Gemma 4 26B A4BGoogle41.2516GB256K24GB GPU or 32GB Macollama pull gemma4:26b-a4b
DeepSeek-R1DeepSeek40.68404GB160K7× 80GB GPUsollama pull deepseek-r1:671b
DeepSeek V3.1DeepSeek40.32404GB160K7× 80GB GPUsollama pull deepseek-v3.1:671b
GLM-4.7-FlashZ.AI38.7419GB198K24GB GPU or 32GB Macollama pull glm-4.7-flash
GPT-OSS 20BOpenAI36.8814GB128K24GB GPU or 24GB Macollama pull gpt-oss:20b
LFM2.5-24B-A2BLiquidAI34.4415GB128K24GB GPU or 24GB Macollama pull lfm2.5:24b
Gemma 4 12BGoogle34.117.7GB256K12GB GPU or 16GB Macollama pull gemma4:12b
Ornith-1.5-9BOrnith AI32.56.6GB256K8GB GPU or 16GB Macollama pull ornith-1.5:9b
Muse Glimmer 30BMeta32.218GB128K24GB GPU or 32GB Macollama pull muse-glimmer:30b
GPT-OSS 120BOpenAI31.9365GB128K80GB GPU or 128GB Macollama pull gpt-oss:120b
DeepSeek V3DeepSeek31.56404GB160K7× 80GB GPUsollama pull deepseek-v3:671b
DeepSeek R1 Distill Qwen 32BDeepSeek30.8820GB128K24GB GPU or 32GB Macollama pull deepseek-r1:32b
Gemma 4 31BGoogle28.8319GB256K24GB GPU or 32GB Macollama pull gemma4:31b
Qwen2.5-72BAlibaba28.5247GB32K80GB GPU or 96GB Macollama pull qwen2.5:72b
Gemma 3 27BGoogle28.4417GB128K24GB GPU or 32GB Macollama pull gemma3:27b
Mistral Medium 3.5 128BMistral28.280GB256K2× 80GB GPUs or 128GB Macollama pull mistral-medium-3.5:128b
Llama 3.1 405BMeta27.98243GB128K4× 80GB GPUs or 512GB Macollama pull llama3.1:405b
Llama 4 MaverickMeta27.68245GB1M4× 80GB GPUs or 512GB Macollama pull llama4:maverick
Llama 4 ScoutMeta26.7567GB10M2× 80GB GPUs or 128GB Macollama pull llama4:scout
Granite 4.2 8BIBM25.925.3GB128K8GB GPU or 16GB Macollama pull granite4.2:8b
Gemma 4 E4BGoogle24.66.6GB128K8GB GPU or 16GB Macollama pull gemma4:e4b
LFM2.5-8B-A1BLiquidAI23.875.2GB125K8GB GPU or 16GB Macollama pull lfm2.5:8b
Gemma 4 E2BGoogle23.764.6GB128K8GB GPU or 16GB Macollama pull gemma4:e2b
Phi-4Microsoft22.899.1GB16K12GB GPU or 16GB Macollama pull phi4:14b
Llama 3 70BMeta20.7240GB8K48GB GPU or 64GB Macollama pull llama3:70b
LFM3-1.7BLiquidAI19.131.1GB256K8GB GPU or 16GB Macollama pull lfm3:1.7b

Each command pulls the size tag we score. Most families also publish other sizes and quantizations (q8_0, bf16, MLX builds for Apple Silicon); see ollama.com/library for the full tag list.

Best Ollama model for coding

On a single 24GB GPU, Qwen3.8-27B has the highest BenchLM coding score you can pull (45.92), from an 18GB download. With server hardware, Ornith-1.5-397B leads the local builds on coding (53.26), but its 242GB download needs 4× 80GB GPUs or a 512GB Mac. Point your editor's OpenAI-compatible endpoint setting at Ollama's local API to use any of them.

Best small Ollama models (8–16GB)

With 8–16GB of VRAM, these are the highest-scoring pulls that leave room for context:

  • Ternary Bonsai 2 27B (51.66): 8.3GB download, fits a 12GB GPU or a 16GB Mac. ollama pull bonsai-2:27b
  • Gemma 4 12B (34.11): 7.7GB download, fits a 12GB GPU or a 16GB Mac. ollama pull gemma4:12b
  • Ornith-1.5-9B (32.5): 6.6GB download, fits an 8GB GPU or a 16GB Mac. ollama pull ornith-1.5:9b
  • Granite 4.2 8B (25.92): 5.3GB download, fits an 8GB GPU or a 16GB Mac. ollama pull granite4.2:8b
  • Gemma 4 E4B (24.6): 6.6GB download, fits an 8GB GPU or a 16GB Mac. ollama pull gemma4:e4b

Open models on Ollama Cloud only

These open-weight models are in the Ollama library only as cloud tags: ollama run sends your prompts to Ollama's hosted service instead of running the weights on your machine, so they are not in the local ranking above. The strongest, Kimi K3.1 Pro (70.37), outscores every local pick.

ModelScoreCloud command
Kimi K3.1 ProMoonshot AI70.37ollama run kimi-k3.1-pro:cloud
Kimi K3Moonshot AI68.45ollama run kimi-k3:cloud
DeepSeek V4.1 Flash 1017DeepSeek64.51ollama run deepseek-v4.1-flash:cloud
DeepSeek V4 Pro 0813DeepSeek62.31ollama run deepseek-v4-pro:cloud
GLM-5.3Z.AI60.98ollama run glm-5.3:cloud
GLM-5.2Z.AI59.74ollama run glm-5.2:cloud
MiMo-V2.6-ProXiaomi57.27ollama run mimo-v2.6-pro:cloud
GLM-5.3-FlashZ.AI56.85ollama run glm-5.3-flash:cloud
Kimi K2.6Moonshot AI56.66ollama run kimi-k2.6:cloud
MiMo-V2.6-FlashXiaomi54.56ollama run mimo-v2.6-flash:cloud
Kimi K2.7 CodeMoonshot AI53.69ollama run kimi-k2.7-code:cloud
Mistral Large 4Mistral52.11ollama run mistral-large-4:cloud
Muse Spark OpenMeta50.73ollama run muse-spark-open:cloud
MiniMax M3MiniMax49.43ollama run minimax-m3:cloud
Nemotron 3 UltraNVIDIA48.74ollama run nemotron-3-ultra:cloud
Beam 1.0 501B-A23BReflection AI43.24ollama run beam:cloud

All Ollama models: the full list with sizes

The Ollama library holds 269 model families as of November 12, 2026. 252 have a build you can run on your own machine, 17 run only on Ollama Cloud, and 14 are embedding models. 41 of them hold a model BenchLM ranks. Pull any local model with ollama pull and its name, or add a size tag such as qwen3.8:27b.

269 of 269 models match.

ModelDefault downloadRuns onAll sizesContextUpdatedScore
qwen3.8-flash-nextvision, tools, thinking120GB125b-a6b-q4_K_M2× 80GB GPUs or 192GB Mac—256KNov 9, 202654.06Qwen3.8-Flash-Next
ripple-previewtools, thinking39GB65b48GB GPU or 64GB Mac—2MNov 9, 2026Not ranked
qwen3.8vision, tools, thinking18GB27b24GB GPU or 32GB Mac—256KNov 8, 202654.06Qwen3.8-27B
autotrust-jevvision, decision11GB9b16GB GPU or 24GB Mac11GB to 26GB, 3 sizes256KNov 4, 2026Not ranked
kimi-k3.1-provision, tools, thinking, cloudCloud onlyOllama Cloud—1MNov 4, 202670.37Kimi K3.1 Pro
mimo-v2.6vision, tools, thinking6.6GB9b8GB GPU or 16GB Mac—256KNov 3, 2026Not ranked
pplx-decider-flashvision, decision5.6GB4b8GB GPU or 16GB Mac—256KNov 3, 2026Not ranked
d3vision, decision5.6GB4b8GB GPU or 16GB Mac3.1GB to 29GB, 4 sizes256KOct 30, 2026Not ranked
qwen3.6vision, tools, thinking23GB35b32GB GPU or 48GB Mac18GB to 23GB, 2 sizes256KOct 30, 202648.05Qwen3.6-27B
lfm3tools, thinking1.1GB1.7b8GB GPU or 16GB Mac—256KOct 29, 202619.13LFM3-1.7B
lfm3-audiovision, tools, audio1.7GB2b8GB GPU or 16GB Mac—125KOct 29, 2026Not ranked
lfm3-omnivision, tools, audio2GB2b8GB GPU or 16GB Mac—125KOct 29, 2026Not ranked
lfm3-vlvision, tools, thinking1.5GB2b8GB GPU or 16GB Mac—125KOct 29, 2026Not ranked
d1vision, decision3.7GB3b8GB GPU or 16GB Mac3.7GB to 3.9GB, 2 sizes32KOct 27, 2026Not ranked
d1-omnivision, decision, audio0.9GB600m8GB GPU or 16GB Mac0.9GB to 4.4GB, 2 sizes125KOct 27, 2026Not ranked
phonon0.1GBde8GB GPU or 16GB Mac—8KOct 24, 2026Not ranked
nemotron-3.5-boltvision, tools, thinking, audio7.2GBe4b12GB GPU or 16GB Mac7.2GB to 25GB, 2 sizes128KOct 23, 202641.61Nemotron 3.5 Bolt 74B-A9B
flinttools, thinking23GB40b32GB GPU or 48GB Mac—256KOct 22, 2026Not ranked
ornith-1.5vision, tools, thinking6.6GB9b8GB GPU or 16GB Mac6.6GB to 242GB, 4 sizes256KOct 20, 202661.41Ornith-1.5-397B
qwen3-vlvision, tools, thinking6.2GB8b8GB GPU or 16GB Mac1.9GB to 143GB, 6 sizes256KOct 20, 2026Not ranked
minimax-m3.1-flashvision, tools, thinking, cloudCloud onlyOllama Cloud—976KOct 19, 2026Not ranked
nemotron-cascade-2tools, thinking18GB30b24GB GPU or 32GB Mac—256KOct 19, 2026Not ranked
nemotron3vision, tools, thinking, audio21GB33b32GB GPU or 48GB Mac—128KOct 18, 2026Not ranked
deepseek-v4.1-flashvision, tools, thinking, cloudCloud onlyOllama Cloud—1MOct 17, 202664.51DeepSeek V4.1 Flash 1017
kevdecision5.6GBlatest8GB GPU or 16GB Mac—256KOct 17, 2026Not ranked
nemotron-3-nanotools, thinking18GB30b24GB GPU or 32GB Mac2.8GB to 18GB, 2 sizes1MOct 17, 2026Not ranked
pplx-decidervision, decision17GB27b24GB GPU or 32GB Mac—256KOct 17, 2026Not ranked
beamtools, thinking, cloudCloud onlyOllama Cloud—1MOct 16, 202643.24Beam 1.0 501B-A23B
bonsai-2vision, tools, thinking3.2GB9b8GB GPU or 16GB Mac0.9GB to 8.3GB, 4 sizes256KOct 16, 202651.66Ternary Bonsai 2 27B
layaembedding, decision0.8GBlatest8GB GPU or 16GB Mac—8KOct 16, 2026Not ranked
nemotron-3.5-lightningtools, thinking19GB30b24GB GPU or 32GB Mac—1MOct 16, 2026Not ranked
embeddinggemma-2vision, embedding, audio1.3GB740m8GB GPU or 16GB Mac0.3GB to 1.3GB, 4 sizes256KOct 15, 2026Not ranked
lfm2.5tools, thinking5.2GB8b8GB GPU or 16GB Mac5.2GB to 15GB, 2 sizes125KOct 15, 202634.44LFM2.5-24B-A2B
muse-spark-openvision, tools, thinking, cloudCloud onlyOllama Cloud—1MOct 15, 202650.73Muse Spark Open
nemotron-3.5-supervision, tools, thinking80GB120b2× 80GB GPUs or 128GB Mac—1MOct 15, 202648.78Nemotron 3.5 Super VL 120B A12B
clefvision, decision, cloud18GB27b24GB GPU or 32GB Mac—256KOct 14, 2026Not ranked
clef-flashvision, decision11GB9b16GB GPU or 24GB Mac—256KOct 14, 2026Not ranked
gemma3nvision9.8GBe4b12GB GPU or 16GB Mac7.8GB to 9.8GB, 2 sizes32KOct 14, 2026Not ranked
glm-ocrvision, tools2.7GBlatest8GB GPU or 16GB Mac—128KOct 14, 2026Not ranked
nimbletools, decision, thinking9.3GB9b12GB GPU or 16GB Mac—256KOct 14, 2026Not ranked

Sizes, context windows, and update dates are Ollama's own figures, read from ollama.com/library on November 12, 2026. Runs on adds about 20% to the download for context and sizes a Mac at three quarters of its unified memory. Quantization and runtime variants of a size (q8_0, MLX builds and the like) are not listed.

How to choose an Ollama model

  • Start from your memory, not the leaderboard. Take the download size and leave 20–30% headroom for context (KV cache). A 24GB card comfortably runs an 18GB download; it will not run a 22GB one at useful context lengths.
  • Prefer MoE models when you must offload to CPU. Sparse mixture-of-experts models (tags like 35b-a3b, where the second number is the active parameter count) compute only 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, sort by the coding category on each model page; the coding leader in a size class is often not the overall leader.
  • Check the license. Open weights don't always mean commercial use is allowed. Each model page links its license, and so does its Ollama library page.
  • Outgrowing one machine? Price a dedicated GPU setup with our self-host cost calculator before renting API capacity.

Questions

What is the best Ollama model right now?

Ornith-1.5-397B is the highest-scoring open-weight model you can pull and run locally, at 61.41 on the November 2026 leaderboard, but its 242GB download needs 4× 80GB GPUs or a 512GB Mac. For a machine you own, Qwen3.8-27B (54.06) is the best answer: an 18GB download that runs on a 24GB GPU or a 32GB Mac.

What is the best Ollama model for coding?

Qwen3.8-27B has the highest BenchLM coding score (45.92) among models that fit a single 24GB GPU. Ornith-1.5-397B leads all local builds on coding (53.26) if you have 4× 80GB GPUs or a 512GB Mac.

How many models does Ollama have?

The Ollama library lists 269 model families as of November 12, 2026. 252 of them have a build you can download and run locally, 17 are available only on Ollama Cloud, and 14 are embedding models. Ollama added or updated 44 families in the 30 days before that date.

How do I list the Ollama models on my machine?

Run ollama ls (ollama list is the same command). It prints each model you have pulled with its size and the date it changed. ollama ps shows the models loaded in memory now, and ollama rm with a model name deletes one. Ollama's CLI reference lists the commands.

How much RAM or VRAM do I need for Ollama?

Take the model's download size and add 20–30% for context. In practice, 8GB runs downloads up to about 6GB (the 7–9B class), 16GB handles about 13GB, and 24GB fits about 20GB, which covers most 27–31B models at Q4_K_M. Macs count unified memory, but macOS gives the GPU only about three quarters of it. Beyond that you are into multi-GPU rigs or high-memory Mac Studios; see our local LLM guide by VRAM tier.

Are Ollama models free?

Local models are. Ollama itself is free, open-source software, and pulling and running an open-weight model costs nothing beyond your hardware and electricity, with no per-token fees. Cloud tags run on Ollama's servers and count against an Ollama account's usage limits. Licenses still vary, so check one before shipping a commercial product on top of a model.

Local LLM updates

Get notified when a new open-weight release changes the Ollama rankings or drops the hardware bar.

Read a sample issue

Join 7,000+ subscribers.

One email each week. Unsubscribe anytime.