Best Ollama Models in 2026 by VRAM: Coding, Chat, Vision and Agents (Ranked)

The best Ollama models in November 2026 are not the ones most “best of” lists still recommend. Llama 3, Mistral 7B and Phi-3 are years old now. In the last three months the Ollama library got Qwen3.8 27B, Gemma 4, Meta’s Muse Glimmer 30B, IBM Granite 4.2, Prism ML’s ternary Bonsai 2, Poolside’s Laguna XS 2.1, Recursive’s Flint coding model, Ornith 1.5 and, just yesterday, Xiaomi’s MiMo-V2.6 9B. This guide picks the best model for each job (coding, chat, vision, agents, tiny devices and big RAM-offload setups) and sorts every pick by how much VRAM it really needs, with the exact ollama pull command for each one.

How I Picked These Models

I went through the whole Ollama library sorted by newest on 5 November 2026 and every model page for the candidates. Sizes in this article are the default download sizes that Ollama shows on each model page (usually Q4_K_M), and benchmark numbers are the ones the model makers publish on those pages. I also checked two independent rankings, BenchLM’s Ollama leaderboard (data as of 4 November 2026) and Morph’s VRAM-ranked list, to make sure I was not missing a popular model.

Three rules decided what made the list:

• It must run locally. Plenty of the biggest names in the library are cloud-only tags. I list those separately so you do not pull one and wonder why your GPU is idle.
• Newer wins when it is clearly better. Qwen3.8 replaces Qwen2.5 and Qwen3 for most jobs, Gemma 4 replaces Gemma 3, and Granite 4.2 replaces Granite 3.x. Older models only stay when they still have a niche.
• Real VRAM, not parameter count. A 27B ternary model can be smaller than a 9B Q8 model, and a 40B mixture-of-experts model can be faster than a 27B dense one. I sort by what you download and what that means on 8, 12, 16 and 24 GB cards.

My own machine has two RTX 5080 cards (16 GB each) and 128 GB of system RAM, so I care a lot about the 16 GB tier and about models that split across two GPUs or spill into RAM. Where I give a speed or fit estimate, it is worked out from the file size and active parameter count, not a stopwatch test, and I say so.

Quick Picks by VRAM

How Much VRAM an Ollama Model Really Needs

The download size is the floor, not the total. On top of the weights, Ollama needs memory for the KV cache (your conversation and context), the vision projector if the model takes images, and a few hundred MB of CUDA overhead. A simple rule that holds up well:

• VRAM needed ≈ download size + 1-2 GB at a 4K-8K context.
• Add roughly 1-4 GB more per 32K tokens of context on a 27B-30B model (it depends on the attention design; hybrid models like Qwen3.5/3.8 and Nemotron need much less).
• If it does not fit, Ollama does not crash. It moves some layers to system RAM and runs them on the CPU. That works, but every layer on the CPU slows generation down.

Mixture-of-experts (MoE) models change the maths. Flint 40B, Qwen3.6 35B, Ornith 1.5 35B and Laguna XS 2.1 all have about 2.5-3B active parameters per token. They need the memory of a 35-40B model, but they generate at roughly the speed of a 3B model, so partial CPU offload hurts them much less than it hurts a dense 27B. That is why a 23 GB MoE can feel faster on a 16 GB card than an 18 GB dense model.

Check what is actually happening with ollama ps. The PROCESSOR column shows 100% GPU when everything fits, or a split such as 23%/77% CPU/GPU when it does not. If you see a split you did not expect, see my Ollama not using GPU fix. For what Q4_K_M, Q5_K_M and Q8_0 mean, read GGUF quantization explained.

Best All-Round Models

1. Qwen3.8 27B: best overall

ollama pull qwen3.8:27b (18 GB, 256K context, text and image input)

Qwen3.8 is the newest dense Qwen model and the one I would install first on any machine with 16 GB of VRAM or more. Ollama lists gains in coding, research and long-horizon agent tasks, native image and video understanding (STEM diagrams, documents and hour-long videos), adjustable thinking with a reasoning_effort setting, and the option to keep earlier reasoning with preserve_thinking. BenchLM ranks it fourth of all local models, just behind Gemma 4 31B and well ahead of every older 27B-32B model.

On a 24 GB card it fits completely. On a 16 GB card a few layers go to the CPU; I would set the context to 16K-32K so the KV cache stays small. There is also a 27b-mlx tag for Apple Silicon.

2. Gemma 4 (12B, 26B and 31B): best for reasoning and maths

ollama pull gemma4:31b (19-20 GB) · gemma4:26b (16-19 GB, MoE, about 3.8B active) · gemma4:12b (7.7-8.0 GB)

Google reports 89.2% on AIME 2026 without tools and a 2150 Codeforces Elo for the 31B, the highest maths and contest-coding numbers of any single-GPU model here. The 12B is the sweet spot for 12 GB cards and gets the full 256K context. The 26B MoE is the fast option on 16-24 GB. All sizes take images, and the small E2B and E4B variants also take audio. Two tips from the model page: put images before your text, and use temperature 1.0, top_p 0.95, top_k 64.

3. Muse Glimmer 30B: best for long agent sessions

ollama pull muse-glimmer:30b (18 GB, 128K context, Apache 2.0)

Meta’s first local model in the Muse family is distilled from Muse Spark and built for “always-on local agents”. Meta’s own table puts it ahead of Gemma 4 31B and Qwen3.6 27B on MCP Atlas (75.5), DeepSearch QA (74.6), SWE-Bench Pro (51.2) and AIME 2026 (94.7), but behind Qwen on SWE-Bench Verified, OSWorld and TerminalBench. If you build tool-calling agents, this is the one to try first. The context is shorter than Qwen’s (128K vs 256K).

4. Granite 4.2 8B: best “boring and reliable” model

ollama pull granite4.2:8b (5.3 GB, 128K) · also :3b (2.2 GB) and :30b (18 GB)

IBM’s Granite 4.2 is Apache 2.0, very good at structured JSON output, summaries, classification and RAG answers, and it has a thinking mode you can switch off with /set nothink. It will not win a coding contest, but for document pipelines and business tools it is one of the most predictable models in the library.

Best Ollama Models for Coding

Flint from Recursive is the surprise of the month. It is a 39.8B MoE with only about 2.5B parameters active per token, Apache 2.0, built for multi-step repository work such as bug hunting, feature work and test writing. Recursive’s own scores are the highest in this table, so treat them as “publisher-reported” until independent tests appear. Because so few parameters are active, it should stay usable on a single 16 GB card with part of the model in RAM, and it fits fully across two 16 GB cards.

Laguna XS 2.1 from Poolside is the other strong local coding agent. Its licence (OpenMDW-1.1) allows commercial use. One warning from the model page: chat on macOS/Metal may return empty output, so use it on Windows or Linux with CUDA, or call /api/generate with "raw": true.

For an 8 GB GPU, MiMo-V2.6 9B is the clear pick. Xiaomi fine-tuned Qwen3.5 9B on data from its much larger MiMo-V2.6 models and lifted SWE Pro from 32.0 to 44.6 and Terminal Bench 2.1 from 27.0 to 37.1. I have a full MiMo-V2.6 9B local setup guide with GGUF sizes and settings.

To use any of these inside VS Code, point Cline, Continue or Roo Code at http://localhost:11434. I walk through that exact setup in my Cline + local LLM guide.

Best Vision Models

Almost every new model in the library now takes images. These are the ones I would use when vision is the main job:

• Qwen3.8 27B (qwen3.8:27b, 18 GB): best overall for documents, charts, screenshots and long video.
• Gemma 4 12B (gemma4:12b, 8 GB): lets you choose a visual token budget (70 to 1120 tokens per image). Use a high budget for OCR and a low one for quick captions.
• Bonsai 2 (bonsai-2:9b 3.2 GB, bonsai-2:27b 8.3 GB): ternary weights (each weight is -1, 0 or +1), about half the size of a normal 4-bit model, Apache 2.0. The 27B in 8.3 GB is remarkable for an 8 GB card.
• Nemotron 3.5 Bolt E4B (nemotron-3.5-bolt:e4b, 7.2-7.9 GB): NVIDIA’s compact VLM built from Gemma 4 E4B, also takes audio. See my Nemotron 3.5 Bolt guide.
• Qwen3-VL 8B (qwen3-vl:8b, 6.1 GB): still a strong OCR model, and it was updated in the library two weeks ago.

For a deeper comparison, including video understanding, see my best local vision and video LLMs roundup.

Best Models for Agents and Tool Calling

An agent model has to call tools in the right format, recover when a tool fails and keep going for dozens of steps. On those jobs, raw intelligence matters less than discipline. My order:

1. Muse Glimmer 30B: best MCP Atlas score (75.5) of any single-GPU model.
2. Qwen3.8 27B: works out of the box with Claude Code, OpenCode, Hermes Agent and OpenClaw according to Ollama’s model page.
3. Flint 40B: for coding agents specifically.
4. Nemotron 3.5 Lightning 30B (nemotron-3.5-lightning:30b, 25 GB, about 3B active, 1M context): NVIDIA’s “always-on” agent model, good when you need a very long context.
5. MiMo-V2.6 9B: the best small agent model; Toolathlon-Verified 35.2 versus 25.9 for its Qwen3.5 base.

The library also gained a new kind of model this autumn: small decision models such as d3, pplx-decider, clef-flash and nimble-v3. They do not chat. They return scores, yes/no answers and labels, which makes them great as cheap routers or judges inside a bigger agent. Use them next to a chat model, not instead of one.

Best Small Models for Laptops and CPUs

On a laptop without a GPU, stay at 4B or below and keep the context at 4K-8K. Liquid AI’s LFM3 family is built for exactly this, and I cover it in my LFM3 local guide.

Big MoE Models for 64-128 GB RAM

If you have a lot of system RAM, you can run models that are far bigger than your GPU. The GPU holds what it can and the CPU runs the rest. Because these are MoE models with a small active size, the speed is often acceptable for chat and slow-but-fine for agents.

• Ornith 1.5 122B (ornith-1.5:122b, 73-77 GB): self-improving model family with vision, tools and thinking. Also comes in 9b (6.6 GB), 35b (23 GB) and 397b (242 GB). BenchLM puts the 397B at the top of all local models.
• Nemotron 3.5 Super 120B (nemotron-3.5-super:120b, about 12B active, 1M context): NVIDIA’s big VLM for coding, documents and tools.
• gpt-oss 120B (gpt-oss:120b, 65 GB, MXFP4): still a strong reasoning model and the cheapest of the three to run.
• Qwen3.8-Flash-Next (125B, 6B active): the preview of the architecture behind Qwen4. In Ollama it is mainly an MLX build for 128 GB Macs; on a PC I run it with llama.cpp instead, as shown in my Blender MCP local LLM guide.

With two GPUs, set OLLAMA_SCHED_SPREAD=1 so Ollama spreads one model across both cards instead of filling one first. On my 2x 16 GB setup that turns 32 GB of VRAM into one pool for a single large model.

Best Embedding Models for RAG

• EmbeddingGemma 2 (embeddinggemma-2:740m, 1.3 GB, also 270m/440m/570m): Google’s new multimodal embedding model on the Gemma 4 architecture. It embeds images and audio as well as text.
• Qwen3 Embedding (qwen3-embedding:0.6b, 4b, 8b): strong multilingual text retrieval.
• nomic-embed-text (274 MB): still the lightweight default in many RAG tools.

Whatever you pick, use the same embedding model for indexing and searching. Switching models means re-embedding every document.

Cloud Tags: What Does Not Run Locally

Ollama now hosts some models on its own servers. Their library pages show a cloud badge and usually no download sizes. Pulling them needs an Ollama account, and your prompts go to Ollama’s cloud, not your GPU. As of today that includes kimi-k3, deepseek-v4.1-flash, deepseek-v4-pro, mistral-large-4, minimax-m3.1-flash, muse-spark-open, glm-5.3, mimo-v2.6-pro and mimo-v2.6-flash.

Several of these have open weights that you can run yourself with llama.cpp and a lot of RAM. See how much VRAM Kimi K3 needs, Mistral Large 4 hardware and running Muse Spark Open locally.

Settings That Make Any Model Faster

These Ollama environment variables are worth setting before you judge a model:

All other commands and variables are in my Ollama commands cheat sheet. If you are still choosing a runner, compare it with LM Studio and llama.cpp in Ollama vs LM Studio vs llama.cpp. When you are done testing, clean up with my guide on how to delete Ollama models, because ten 20 GB models fill a drive fast.

Frequently Asked Questions

What is the best Ollama model in 2026?

For most people with 16-24 GB of VRAM, Qwen3.8 27B (ollama pull qwen3.8:27b) is the best all-round local model. Gemma 4 31B is better at maths and reasoning, and Muse Glimmer 30B is better for long agent sessions.

What is the best Ollama model for coding?

Flint 40B (23 GB, about 2.5B active) has the highest publisher-reported coding scores of any local model in the library. Laguna XS 2.1 and Qwen3.6 35B are close. On an 8 GB GPU, use MiMo-V2.6 9B.

What is the best Ollama model for 8 GB VRAM?

MiMo-V2.6 9B for coding and agents, Gemma 4 12B for chat and vision, and Bonsai 2 27B if you want a 27B model in an 8.3 GB file.

What is the best Ollama model for 16 GB VRAM?

Qwen3.8 27B with a 16K-32K context, accepting a small CPU offload. If you want everything on the GPU, use Gemma 4 12B or Gemma 4 26B, which is a fast MoE.

Which Ollama models run without a GPU?

Bonsai 2 2B and 4B, LFM3 1.7B, Qwen3.5 0.8B to 4B, Gemma 4 E2B and Granite 4.2 3B all run acceptably on a modern CPU with 8-16 GB of RAM.

Why does my Ollama model run on the CPU?

Either the model plus its context does not fit in VRAM, or Ollama cannot see your GPU. Run ollama ps to see the CPU/GPU split, lower the context length, or pick a smaller model.

Are cloud models in Ollama free and private?

Cloud tags run on Ollama’s servers, need an account and are subject to Ollama’s limits and terms. They are not processed on your own machine, so do not use them for data that must stay local.

Final Thoughts

If you only install three models this month, make them Qwen3.8 27B for everyday work, MiMo-V2.6 9B as a fast small coder and Flint 40B if you have the memory for a proper local coding agent. Add Gemma 4 12B on smaller cards and Bonsai 2 on laptops. The library is moving fast (nine of the models above did not exist three months ago), so I will keep this list updated as new tags land.

New to Ollama? Start with how to install Ollama, then come back here to pick a model. For uncensored options see the best uncensored local LLMs.

Sources

• Ollama library (newest) and model pages for qwen3.8, gemma4, muse-glimmer, flint, laguna-xs-2.1, mimo-v2.6, bonsai-2, ornith-1.5, granite4.2, nemotron-3.5-bolt, qwen3.6, qwen3.5
• BenchLM: Best Ollama Models (data as of 4 Nov 2026)
• Morph: Best Ollama Models ranked by VRAM
• XiaomiMiMo/MiMo-V2.6-Distill-Qwen-9B model card
• Ollama FAQ