Skip to main content
PromptQuorumBuilt for humans. Structured for AI.
Home/Local LLMs
Best Models

Ollama Latest Version 2026: v0.40.5 + Best Models by Use Case

·9 min readBy · Founder of PromptQuorum · Discovery engine for open-weight & open-source AI

The current stable Ollama version is v0.40.5 (October 14, 2026); v0.40.6 (October 16) is a pre-release. The newest models added recently are NVIDIA Nemotron 3.5 Super (120B, vision-language) and Bolt (74B total / 9B active MoE, 25 GB) in v0.40.5, with Nemotron 3.5 Lightning (30B MoE, 3B active) for always-on agents. Best overall on consumer hardware is Qwen3.8-27B (61.7 SWE-bench Pro as reported by Qwen, 18 GB, fits a 24 GB card). Other top picks: Devstral Small 2 and Laguna XS 2.1 (local agentic coding; Kimi K2.6 is frontier-scale but cloud-only on Ollama), gpt-oss:20b (best small / 16 GB), qwen3:30b (balanced all-round), DeepSeek-R1 (reasoning), Gemma 4 (vision/tool calling, now also 12B and 31B), Granite 4.2 (IBM, Apache 2.0, 3B/8B/30B), Dolphin 3.0 (uncensored general-purpose), and Llama 4 Scout (long-context 10M / multimodal). Most downloaded overall remains the Llama family.

Key Takeaways

  • Best overall on consumer hardware: Qwen3.8-27B (61.7 SWE-bench Pro as reported by Qwen, 18 GB download, fits a 24 GB card). Balanced all-round: qwen3:30b.
  • Most downloaded: Llama 3.2 3B (tutorials) and the Llama family -- widest tool support.
  • Best reasoning: DeepSeek-R1 (chain-of-thought) and gpt-oss:20b (adjustable reasoning, ~o3-mini level).
  • Best coding: Qwen3.8-27B (best dense), Laguna XS 2.1 (agentic, long-horizon), Devstral Small 2 (best agentic 24B), qwen3-coder:30b (completion); Kimi K2.6 is frontier-scale but cloud-only on Ollama -- highest reported benchmarks at their sizes.
  • Best small / 16 GB: gpt-oss:20b. Best vision/multimodal: Gemma 4 (E4B+). Best long-context (10M) / large multimodal: Llama 4 Scout (~55 GB). Best uncensored/general-purpose: Dolphin 3.0. Always-on agents: Nemotron 3.5 Lightning.
  • The Ollama library holds several hundred curated models (thousands more via custom GGUF Modelfiles). All are available via `ollama pull <name>`.

📍 In One Sentence

The best Ollama model in 2026 is Qwen3.8-27B (61.7 SWE-bench Pro as reported by Qwen, 18 GB); best local coding picks are Qwen3.8-27B, Devstral Small 2 and Laguna XS 2.1 (Kimi K2.6 is cloud-only on Ollama).

💬 In Plain Terms

Ollama is a free tool that lets you run AI models on your own computer — no internet or API key needed. These are the top models you can download with a single command, organized by what you actually need them for.

The current stable Ollama version is v0.40.5 (October 14, 2026), which adds NVIDIA Nemotron 3.5 Super and Bolt, decision-model support and updated llama.cpp and MLX engines; v0.40.6 is a pre-release. Since v0.33.1 in August, Ollama made MLX the default engine on Apple Silicon (v0.40.0) and added ChatGPT Desktop integration (v0.34.0). Best overall on consumer hardware is still Qwen3.8-27B (61.7 SWE-bench Pro as reported by Qwen, 18 GB download, fits a 24 GB card or 32 GB Mac); best for coding depends on your setup — see the use-case breakdown below.

Slide Deck: Ollama Latest Version 2026: v0.40.5 + Best Models by Use Case

The slide deck below covers: top Ollama models by download count, performance comparison, best models by use case (chat, coding, agentic coding, reasoning, vision), DeepSeek-R1 chain-of-thought reasoning, and exact pull commands. Download the PDF as your Ollama model selection reference card.

Browse the slides below or download as PDF for offline reference. Download Reference Card (PDF)

Ollama Latest Version 2026: v0.40.5 + Best Models by Use Case

What Is the Latest Ollama Version and What Changed?

Current Ollama version: v0.40.5 (released October 14, 2026). This is the latest stable release, available via ollama.com/download. A v0.40.6 pre-release landed October 16 but is not yet marked stable. Update with `curl https://ollama.ai/install.sh | sh` (macOS: `brew upgrade ollama`), then confirm with `ollama --version`.

What changed in the v0.33–v0.40 series (August–October 2026): v0.33.0 (August 21) enabled Claude Desktop as a third-party gateway provider and made KV-cache prefill restore points more reliable. v0.34.0 (September 5) let ChatGPT Desktop use Ollama models and sped up structured output on Apple Silicon; v0.34.1 (September 14) made MLX safetensors `ollama create` non-experimental. v0.40.0 (September 25) runs supported architectures such as qwen3.8, gemma4 and qwen3.6 on MLX by default on Apple Silicon. v0.35.0 (September 28) added decision models through `/v1/systemone` — models that return choices, probabilities and scores instead of text, starting with Nimble and Tev1 — and v0.35.1 (September 29) added the multimodal Clef models. v0.40.2 (October 8) keeps backups while upgrading older models in the background, v0.40.4 (October 12) added `ollama run <decision-model> --ask`, and v0.40.5 (October 14) added NVIDIA Nemotron 3.5 Super support plus cloud access for Clef decision models. Full notes: github.com/ollama/ollama/releases.

Newest models in the library (2026):

  • Nemotron 3.5 Lightning, Super and Bolt (NVIDIA, August–October 2026) — Lightning (August 11) is a 30B MoE with 3B active parameters for always-on agents on a single GPU; Super is a 120B vision-language MoE (12B active, up to 1M-token context, tools and thinking mode); Bolt distills Super to 74B (9B active) to fit a 24 GB NVIDIA GPU. Pull: `ollama pull nemotron-3.5-lightning`, `ollama run nemotron-3.5-super`, `ollama run nemotron-3.5-bolt`
  • Laguna XS 2.1 (Poolside, July 2, 2026) — 33B total / 3B active MoE built for agentic coding and long-horizon local work: plan, write code, run tests, iterate. 256K context, OpenMDW-1.1 license. SWE-bench Verified 70.9%, SWE-bench Multilingual 63.1%. Pull: `ollama pull laguna-xs-2.1`
  • Kimi K2.7 Code (Moonshot AI, June 2026) — Coding-focused agentic model built on Kimi K2.6, aimed specifically at long-horizon coding sessions. Pull: `ollama pull kimi-k2.7-code:cloud` (cloud tag; local GGUF builds need 300+ GB)
  • DeepSeek V4 Pro (DeepSeek, April 23, 2026) — Algorithmic-coding specialist, 93.5% LiveCodeBench, MIT license. Budget sibling DeepSeek V4 Flash for lighter hardware. Pull: `ollama pull deepseek-v4-pro`
  • Kimi K2.6 (Moonshot AI, April 20, 2026) — Frontier coding model, SWE-Bench Pro 58.6, SWE-bench Verified 80.2%. MoE architecture (32B active / 1T total). Modified MIT license. Pull: `ollama pull kimi-k2.6:cloud` (cloud tag; local GGUF builds need 300+ GB)
  • Qwen3.8-27B (Alibaba, April 16, 2026) — Best overall on consumer hardware, 61.7 SWE-bench Pro, Apache 2.0, fits 24 GB at Q4. Also Qwen3.6-35B-A3B (MoE, 73.4 SWE-bench). Pull: `ollama pull qwen3.8:27b`
  • GLM-5.3 (Z.ai) — 744B / 40B active MoE, MIT license. Structured code generation leader. Pull: `ollama pull glm-5.3`
  • gpt-oss (OpenAI, 2026) — Open-weight MoE: gpt-oss:20b (21B total / 3.6B active, runs in 16 GB, ~o3-mini level, adjustable reasoning) and gpt-oss:120b (80 GB). Pull: `ollama pull gpt-oss:20b`
  • Gemma 4 (Google, April 2, 2026) — Multimodal sizes E2B / E4B / E12B (26B MoE) / E27B (31B dense), all with vision and tool calling. Now ~90% faster on Apple Silicon (July 2026 MLX update). E4B runs in ~6 GB VRAM. Pull: `ollama pull gemma4:e4b`
bash
# Update Ollama to the latest version (v0.40.5)
curl https://ollama.ai/install.sh | sh

# Or on Mac: brew upgrade ollama

# Check your current version
ollama --version  # outputs: ollama version 0.40.5

# Pull the newest 2026 models
ollama pull laguna-xs-2.1
ollama pull kimi-k2.7-code:cloud
ollama pull nemotron-3.5-lightning
ollama pull kimi-k2.6:cloud

What Is the Ollama Version History and Release Timeline?

Ollama published its first release, v0.0.1, on July 8, 2023, and reached v0.40.5 on October 14, 2026. The table lists the releases that changed what you can do with Ollama; every tag, with full notes, is at github.com/ollama/ollama/releases. Version numbers jumped from v0.35 to v0.40 in late September 2026, so v0.40.0 (September 25) sits next to the v0.35 releases (September 28-29) on GitHub.

Version
Released
Key change
v0.40.5Oct 14, 2026Nemotron 3.5 Super; cloud decision models
v0.40.0Sep 25, 2026MLX by default on Apple Silicon
v0.35.0Sep 28, 2026Decision models via /v1/systemone
v0.34.0Sep 5, 2026Ollama models in ChatGPT Desktop
v0.33.0Aug 21, 2026Claude Desktop gateway; KV-cache restore
v0.32.0Jul 11, 2026"Chat, Code & Work" agent experience
v0.0.1Jul 8, 2023First public release

Which Ollama Models Work Best for Your Use Case?

The quality of a model's output depends heavily on how you prompt it. For structured techniques that work across all local models — including chain-of-thought, few-shot examples, and output formatting — see the prompt engineering guide. For reasoning tasks, chain-of-thought prompting significantly improves DeepSeek-R1 and Qwen3 output quality. To understand quantization tradeoffs for these models, see the quantization guide →. For determining how much VRAM each model needs, see the VRAM requirements guide →. For agent workflows with Gemma 4, see Tree-of-Thought and ReAct. For hardware requirements to run these models, see the hardware guide →. Once a tool-calling model from this list is wired into a multi-step loop with file and database access, see Local AI Agents With MCP for the open-source orchestration pattern.

📍 In One Sentence

For general chat use Qwen3.8-27B, for coding use Qwen3.8-27B, Devstral Small 2 or Laguna XS 2.1, for reasoning use DeepSeek-R1, for vision use Gemma 4 E4B, for uncensored chat use Dolphin 3.0.

💬 In Plain Terms

Different AI models excel at different tasks — like how a calculator beats a word processor at math. This section matches the right model to each job.

  • General chat (beginner): `ollama run llama3.2:3b` -- most documentation, best-supported first model.
  • General chat (best overall): `ollama run qwen3.8:27b` -- 61.7 SWE-bench Pro, best overall on consumer hardware, fits 24 GB at Q4. Balanced all-round: `ollama run qwen3:30b`. For 8 GB machines, keep `ollama run llama3.2:3b`.
  • Long-context / multimodal: `ollama run llama4:scout` -- 10M-token context + multimodal, MoE (17B active/109B total). Needs ~55 GB VRAM at Q4 (fits 24 GB only at 1.78-bit ~20 tok/s).
  • Best small / 16 GB: `ollama run gpt-oss:20b` -- 21B total / 3.6B active MoE, ~o3-mini level, adjustable reasoning. Larger: `ollama run gpt-oss:120b` (80 GB).
  • Coding on 8 GB: `ollama run qwen3:8b` -- Strong coding option for 8 GB VRAM machines. 5 GB used, multilingual.
  • General inference on 8 GB (if not coding): `ollama run mistral:7b` -- Lightweight general-purpose model for 8 GB machines; speed depends on your hardware.
  • Coding (best agentic, 24B): `ollama run devstral-small-2` -- Best agentic coding model (multi-file edits, debugging). 16 GB RAM. By Mistral AI.
  • Coding (best dense, 27B): `ollama run qwen3.8:27b` -- 61.7 SWE-bench Pro (Qwen-reported). Best dense coding model. 18 GB download.
  • Coding (frontier MoE): `ollama run kimi-k2.6:cloud` -- SWE-Bench Pro 58.6 (ties GPT-5.5), top tier. MoE (32B active/1T total). Modified MIT license. Ollama lists only a cloud tag; local GGUF builds need 300+ GB.
  • Coding (agentic, long-horizon): `ollama run laguna-xs-2.1` -- Poolside's 33B/3B-active MoE, SWE-bench Verified 70.9%, 256K context. Built for multi-step plan → code → test → iterate loops. OpenMDW-1.1 license.
  • Agent tasks and tool calling: `ollama run gemma4:e4b` -- Released April 2, 2026. Built-in tool calling + vision support. Recommended for local agents, function calling, and structured output. 6 GB RAM.
  • Reasoning and math: `ollama run deepseek-r1:7b` -- chain-of-thought model, best local math performance at 7B.
  • Multilingual (general): `ollama run qwen3:8b` -- 29+ native languages, a multilingual general-purpose option.
  • Chinese-language tasks: `ollama run qwen3.8:27b` -- native Chinese tokenization, Alibaba origin. See Regional Context for CAC compliance notes.
  • Japanese-language tasks: `ollama run qwen3.8:27b` -- multilingual model that lists Japanese among its supported languages; compare output on your own text.
  • Russian-language tasks: `ollama run qwen3:8b` or `ollama run mistral-small3.1` -- both include Russian in their native multilingual training data. This page does not rank the two for Russian, so test both on your own prompts.
  • Uncensored / general-purpose: `ollama run dolphin3` -- Dolphin 3.0 (Cognitive Computations, built on Llama 3.1), no built-in content filtering, general chat/coding/agentic use.
  • Image understanding: `ollama run gemma4:e4b` -- vision + tool calling. Or `ollama run llama3.2-vision:11b` for dedicated vision.
  • Fast and lightweight: `ollama run gemma2:2b` -- fastest CPU inference, 1.7 GB RAM.
  • High quality (16 GB RAM): `ollama run mistral-small3.1` -- strong quality at 14 GB RAM.
  • Embedding generation: `ollama run nomic-embed-text` -- 137M parameter embedding model for RAG pipelines.
  • Document Q&A (RAG): `ollama run llama3.2` with Open WebUI's RAG feature -- best-supported combination.
  • Home automation / wake word AI: `ollama run phi4-mini` — Phi-4 Mini (3.8B, ~3 GB VRAM) can handle Home Assistant voice queries on a mini PC without a discrete GPU; speed depends on your hardware. See Home Assistant + Ollama integration guide →.
Ollama model selection by use case: pick qwen3.8:27b (best overall, 61.7 SWE-bench Pro) for chat and coding, kimi-k2.6 for frontier coding, gpt-oss:20b on 16 GB, deepseek-r1:7b for math.
Ollama model selection by use case: pick qwen3.8:27b (best overall, 61.7 SWE-bench Pro) for chat and coding, kimi-k2.6 for frontier coding, gpt-oss:20b on 16 GB, deepseek-r1:7b for math.

Which Models Were Added to Ollama Most Recently?

These are the newest models in the Ollama library, newest first. Confirm availability with `ollama pull <model>` before building workflows — new models appear at ollama.com/library within days of release.

Model
Released
Best For
Ollama Command
nemotron-3.5-superOctober 2026NVIDIA — vision-language MoE, 120B/12B active, 1M context, toolsollama run nemotron-3.5-super
nemotron-3.5-boltOctober 2026NVIDIA — 74B/9B active MoE sized for a 24 GB GPUollama run nemotron-3.5-bolt
nemotron-3.5-lightningAugust 11, 2026NVIDIA — 30B/3B active MoE for always-on agentsollama run nemotron-3.5-lightning
laguna-xs-2.1July 2, 2026Poolside — agentic coding, 33B/3B active MoE, SWE-bench Verified 70.9%, 256K contextollama run laguna-xs-2.1
kimi-k2.7-codeJune 2026Moonshot AI — coding-focused agentic model built on Kimi K2.6ollama run kimi-k2.7-code:cloud
deepseek-v4-proApril 23, 2026Algorithmic coding, 93.5% LiveCodeBench, MITollama run deepseek-v4-pro
kimi-k2.6April 20, 2026Frontier coding (SWE-Bench Pro 58.6), MoE (32B/1T), Modified MITollama run kimi-k2.6:cloud
qwen3.8:27bApril 16, 2026Best overall on consumer hardware, 61.7 SWE-bench Pro, fits 24 GB Q4ollama run qwen3.8:27b
qwen3:30b2026Balanced all-round; qwen3-coder:30b for code completionollama run qwen3:30b
gpt-oss:20b2026Best small / 16 GB, ~o3-mini, adjustable reasoning (also gpt-oss:120b)ollama run gpt-oss:20b
glm-5.3Z.ai744B/40B active MoE, MIT licenseollama run glm-5.3
gemma4:e4bApril 2, 2026Vision + tool calling (E2B/E4B/E12B/E27B)ollama run gemma4:e4b
deepseek-v4-flashApril/May 2026Budget coding optionollama run deepseek-v4-flash
qwen3:8b20267B-class multilingual modelollama run qwen3:8b

What Is DeepSeek-R1 and Why Is It Different?

DeepSeek-R1 is a reasoning model -- unlike standard chat models that generate answers directly, DeepSeek-R1 generates explicit chain-of-thought reasoning before its final answer. This significantly improves performance on math, logic puzzles, and step-by-step problem solving.

DeepSeek-R1 7B is a chain-of-thought reasoning model. It is slower than standard models (more tokens per response) but significantly more accurate on tasks where reasoning matters.

Version note: DeepSeek-R1 is a previous-generation model. DeepSeek's current open-weight family is DeepSeek-V4, which the Ollama library did not list when checked, so the deepseek-r1 distills (1.5B to 70B) remain the reasoning option you can pull there. DeepSeek has replaced the hosted R1 API model with DeepSeek-V4.

bash
# Pull and run DeepSeek-R1
ollama run deepseek-r1:7b

# Larger variants for better quality
ollama run deepseek-r1:14b   # 10 GB RAM
ollama run deepseek-r1:32b   # 20 GB RAM

Which Ollama Models Support Image Input?

These models on Ollama support image input (multimodal): Gemma 4 supports both vision AND tool calling — unique among vision models on Ollama, and now runs nearly 90% faster on Apple Silicon after the July MLX update.

Model
RAM
Image Support
Ollama Command
llama3.2-vision:11b~8 GBYesollama run llama3.2-vision:11b
llama3.2-vision:90b~55 GBYesollama run llama3.2-vision:90b
gemma3:9b (vision)~6 GBYesollama run gemma3:9b
minicpm-v:8b~5.5 GBYesollama run minicpm-v
gemma4:e4b~6 GBYes + Tool Calling ✓ollama run gemma4:e4b
5 Ollama vision models for image input. Gemma 4 E4B (6 GB) now includes tool calling. Llama 3.2 Vision 11B (8 GB) for dedicated vision. All run locally.
5 Ollama vision models for image input. Gemma 4 E4B (6 GB) now includes tool calling. Llama 3.2 Vision 11B (8 GB) for dedicated vision. All run locally.

What Are the Top 10 Open Source Models on Ollama?

Download counts still favor Llama 3.x due to tutorial prevalence. For new projects in 2026, prefer Qwen3.8-27B (best overall on consumer hardware), Laguna XS 2.1 or Devstral Small 2 (coding), gpt-oss:20b, and qwen3:30b.

#
Model
Best For
RAM
HumanEval
1Qwen3.8-27BBest overall on consumer hardware24 GB (Q4)61.7 SWE-bench Pro
2Kimi K2.6Frontier coding, MoE (32B/1T), Modified MITQuantized58.6 SWE-Bench Pro
3gpt-oss:20bBest small / 16 GB, adjustable reasoning16 GB~o3-mini
4qwen3:30bBalanced all-round; qwen3-coder:30b for code~18 GBstrong
5Laguna XS 2.1Agentic coding, long-horizon (plan/code/test/iterate)Quantized70.9% SWE-bench Verified
6Devstral Small 2 24BAgentic coding (multi-file)15 GB68.0% (SWE-bench Verified)
7deepseek-r1:7bReasoning, math5 GB—
8gemma4:e4bVision + tool calling (multimodal)~6 GB—
9Llama 4 ScoutLong-context 10M + multimodal, MoE~55 GB (Q4)—
10Llama 3.2 3BFirst model, general chat2.5 GB—
Top Ollama models 2026: Qwen3.8-27B (best overall, 24 GB Q4), Kimi K2.6, Laguna XS 2.1 (agentic coding), gpt-oss:20b. Llama 4 Scout for 10M-token context (~55 GB).
Top Ollama models 2026: Qwen3.8-27B (best overall, 24 GB Q4), Kimi K2.6, Laguna XS 2.1 (agentic coding), gpt-oss:20b. Llama 4 Scout for 10M-token context (~55 GB).

How Do You Browse the Ollama Model Library?

There are two ways to work with Ollama models. Switch installed models: In the Ollama Mac app, click the model dropdown button at the bottom of the chat input (shows the current model name, e.g. "gemma3:1b") to switch between any locally installed model. Find and download new models: Visit ollama.com/library to browse several hundred curated models by category, then use the CLI commands below to pull and manage them.

bash
# List all locally downloaded models
ollama list

# Search for a model and pull it
ollama pull qwen2.5-coder:32b

# See all available tags for a model
ollama show qwen2.5

# Remove a model to free disk space
ollama rm llama3.2:3b

How Do Regional Privacy Rules Affect Your Ollama Model Choice?

EU / GDPR + Licence Compliance. For EU organizations deploying Ollama models in production, licence choice matters as much as performance. Apache 2.0 (fully open, commercial use permitted): Mistral Small 3, Mistral Small 3.1, Qwen3 8B, Qwen3.8-27B, Devstral Small 2, Granite 4.2. Meta Llama Community Licence (commercial use restricted above 700M monthly active users): Llama 3.1 8B, Llama 3.2 3B, Llama 3.2 Vision 11B. Llama 3.2 Vision is an exception: its Acceptable Use Policy does not grant the licence rights to individuals or companies domiciled in the EU (end users of products built on it are not affected), so check it before use in the EU. MIT (commercial use permitted): DeepSeek-R1 7B, DeepSeek-R1 14B. Modified MIT (commercial use permitted with attribution clause): Kimi K2.6. OpenMDW-1.1 (permissive, commercial use permitted): Laguna XS 2.1. Always check the licence terms before commercial deployment. For EU enterprises in regulated sectors, Mistral models (France, Apache 2.0) such as Devstral Small 2 are a sensible default -- EU origin and a permissive licence. Running models entirely on-premises via Ollama means prompts are not sent to external servers; this supports, but does not by itself satisfy, GDPR obligations (a `:cloud` tag does send data to Ollama's hosted service).

Japan (METI). For Japanese enterprise Ollama deployments, Qwen3 / Qwen3.8 is a recommended model family -- it handles Japanese text natively, which helps inference time and KV cache requirements. For Japanese coding workflows: Qwen3.8-27B (61.7 SWE-bench Pro as reported by Qwen) handles Japanese code comments natively. Recording the exact model version is good practice for compliance documentation. Use `ollama show <model>` to get the full model specification including parameter count, quantization level, and context length for compliance records.

China. China's CAC Generative AI Measures (2023) can impose filing or security-assessment duties on providers of generative AI services to the public in China; whether they apply to you is a question for your legal adviser. Qwen3 / Qwen3.8 (Alibaba, Apache 2.0) is a natural choice for Chinese enterprise Ollama deployments -- Chinese model origin, Apache 2.0 licence, strong Chinese-language performance. Kimi K2.6 (Moonshot AI, Modified MIT) is also Chinese-origin, but Ollama lists it only as a cloud tag and local runs need server-class hardware. Pull commands: `ollama run qwen3.8:27b` for best quality, `ollama run qwen3:8b` for speed. DeepSeek-R1 (MIT licence) is appropriate for reasoning tasks. Inference that runs locally via Ollama does not by itself send data abroad, which removes the cross-border transfer question for that step; whether PIPL or other obligations apply to your processing is for you and your legal adviser to decide. This section is general information, not legal advice.

What Are the Common Mistakes When Choosing Ollama Models?

Pulling the largest model tag by default without checking RAM

Running `ollama pull llama3.3` without specifying a tag downloads the default variant, which is typically the largest standard quantization. On a machine with 8 GB RAM, pulling llama3.3 (70B at ~40 GB) will fail or cause severe swap usage. Always specify the variant: `ollama pull llama3.2:3b` for 8 GB machines.

Using a general model when a task-specific model exists

For coding tasks, `qwen2.5-coder:7b` is built for code and includes FIM (fill-in-the-middle) support for code completion, which the general `qwen2.5:7b` does not. For reasoning/math, `deepseek-r1:7b` is a dedicated chain-of-thought reasoning model. Task-specific models exist in the Ollama library for a reason.

Not verifying a model is available before building a workflow

The Ollama library changes over time -- models are added and occasionally removed. Before building a production pipeline around a specific model, confirm it is in the library (`ollama list` locally, or check ollama.com/library). Pin specific model versions in production workflows: `ollama pull llama3.1:8b-instruct-q4_K_M`.

Not specifying a quantization tag for large models

Running `ollama pull qwen2.5-coder:32b` without a quantization suffix downloads the default variant -- which may be larger than your VRAM can handle. For 16 GB VRAM, pull the explicit Q4_K_M variant: `ollama pull qwen2.5-coder:32b-instruct-q4_K_M`. Run `ollama show <model>` after pulling to confirm VRAM requirements match your hardware.

Expecting DeepSeek-R1 to be as fast as standard chat models

DeepSeek-R1 generates explicit chain-of-thought reasoning tokens before its final answer -- this is why it outperforms standard models on math and logic, but it produces 3-5x more tokens per response. For quick chat or one-line answers, use `llama3.1:8b`. Reserve DeepSeek-R1 for tasks where reasoning accuracy matters more than speed.

Next steps

Frequently Asked Questions

How many models are in the Ollama library?

The Ollama library holds several hundred curated models (official support), growing weekly as new open-weight releases land. Hugging Face hosts thousands of additional GGUF models that can be loaded via Ollama using custom Modelfiles.

What is the official name of the company behind Ollama?

The company behind Ollama is Ollama Inc. "Ollama" refers to both the company and the open-source CLI tool and model runtime it publishes; the project's first public release, v0.0.1, appeared on GitHub on July 8, 2023.

Can I use models from Hugging Face directly in Ollama?

Yes. Download a GGUF file from Hugging Face and create a Modelfile: `FROM ./model.gguf`. Then run `ollama create mymodel -f Modelfile`. This works for any GGUF file including fine-tunes and models not in the official Ollama library.

Which Ollama model is best for building a local chatbot?

For a general-purpose local chatbot: `qwen3.8:27b` (best overall on consumer hardware, fits 24 GB at Q4), or `llama3.2:3b` on 8 GB RAM (easiest entry point). For 16 GB machines: `gpt-oss:20b` (~o3-mini level) or `mistral-small3.1`. For a coding assistant chatbot: `qwen3.8:27b` (61.7 SWE-bench Pro), `devstral-small-2` (15 GB, agentic), or `laguna-xs-2.1` (agentic, long-horizon). Pair with Open WebUI for a web-based interface that connects to Ollama's API at localhost:11434.

Are all Ollama models truly open source?

Not all. The Ollama library includes models with varying licences. Llama 3.x/4.x use the Meta Llama Community Licence (not OSI-approved open source -- restricts commercial use above 700M monthly active users). Mistral Small 3.1, Qwen3, Qwen3.8, Devstral Small 2 and Gemma 4 are Apache 2.0 (fully open source); Gemma 2 and Gemma 3 use Google's Gemma Terms of Use, which are not OSI-approved. Kimi K2.6 is Modified MIT licensed (commercial-friendly with an attribution clause). Laguna XS 2.1 uses OpenMDW-1.1 (permissive, commercial use permitted, not OSI-approved). Always check the licence before commercial deployment.

Which Dolphin model should I use with Ollama?

Use Dolphin 3.0 (`ollama pull dolphin3`), not the older `dolphin-mistral` (last updated 2024, built on Mistral 0.2). Dolphin 3.0, maintained by Cognitive Computations (Eric Hartford), is built on Llama 3.1, has no built-in content filtering, and is designed as a general-purpose local model for coding, math, agentic tasks, and chat without built-in content restrictions. Because it ships without built-in safety filtering, you are responsible for how its output is used, and the Llama 3.1 licence terms apply.

Which embedding model should I use with Ollama for RAG?

`nomic-embed-text` is the standard choice -- a 137M parameter model that generates 768-dimensional embeddings, runs at milliseconds per document, and is specifically designed for retrieval tasks. Pull it with `ollama pull nomic-embed-text`. Use with Open WebUI's built-in RAG, LangChain's OllamaEmbeddings, or LlamaIndex.

How often does the Ollama library get updated with new models?

The Ollama team adds new models within days to weeks of major releases. Laguna XS 2.1 (July 2, 2026), Kimi K2.7 Code, Kimi K2.6 and Qwen3.8 all appeared within days of their releases. The current Ollama version is v0.40.5 (October 14, 2026). Follow the Ollama GitHub repository (github.com/ollama/ollama) or the Ollama Twitter/X account for new model announcements.

What is the difference between `ollama pull` and `ollama run`?

`ollama pull` downloads the model file to local storage (one-time operation). `ollama run` starts an interactive session immediately after pulling, or reuses the already-pulled model if available. You can pull once and run multiple times without re-downloading.

Can I run multiple models simultaneously on the same machine?

Yes, if your hardware has sufficient VRAM. Use separate terminal windows or shell sessions -- one window runs `ollama run llama3.2` while another runs `ollama run qwen2.5:7b`. Ollama automatically manages VRAM sharing. Monitor `nvidia-smi` or system activity to avoid overload.

How do I update a model to the latest version?

`ollama pull [model-name]` checks for updates and downloads the latest version if available. To revert or use specific versions, use version tags: `ollama pull llama3.1:8b` or `ollama pull llama3.1:8b-instruct-q4_K_M`. Check available versions with `ollama show [model-name]`.

Are open source models on Ollama truly free to use commercially?

Most are, but not all. Llama 3.x (Meta Llama Community Licence) restricts commercial use above 700M monthly active users. Mistral Small 3.1, Qwen3 and Gemma 4 use Apache 2.0 (fully commercial-friendly); Gemma 2 and Gemma 3 use Google's Gemma Terms of Use, which permit commercial use under conditions. Always verify the licence before enterprise deployment -- check the model's Hugging Face page or Ollama library entry.

What is the latest version of Ollama?

The latest stable Ollama version is v0.40.5, released October 14, 2026. A v0.40.6 pre-release followed on October 16 but is not yet marked stable. Download it from ollama.com/download, or update in place with `curl https://ollama.ai/install.sh | sh` on Linux and `brew upgrade ollama` on macOS.

Where can I find the Ollama changelog and release notes?

Full release notes for every version live at github.com/ollama/ollama/releases, which lists each tag with its date and changelog. The most recent stable entries are v0.40.5 (October 14, 2026 -- NVIDIA Nemotron 3.5 Super support, cloud access for decision models), v0.40.4 (October 12 -- decision models in `ollama run --ask`, faster loading) and v0.40.0 (September 25 -- MLX by default on Apple Silicon).

How do I check which Ollama version I have installed?

Run `ollama --version` in your terminal. It prints a line such as `ollama version 0.40.5`. If the number is lower than the current release, update with `curl https://ollama.ai/install.sh | sh` (Linux) or `brew upgrade ollama` (macOS), then run `ollama --version` again to confirm.

When was Ollama first released?

Ollama's first release, v0.0.1, was published on GitHub on July 8, 2023, shortly after the repository was created on June 26, 2023. By October 2026 the project had reached v0.40.5. The full release list is at github.com/ollama/ollama/releases.

What are the newest Ollama models in 2026?

The latest additions are NVIDIA's Nemotron 3.5 family (Lightning: 30B MoE, 3B active; Super: 120B vision-language MoE, 12B active, up to 1M-token context; Bolt: 74B, 9B active, sized for a 24 GB GPU), alongside Laguna XS 2.1 (Poolside, OpenMDW-1.1 -- agentic long-horizon coding, SWE-bench Verified 70.9%, 33B/3B active MoE) and Kimi K2.7 Code (Moonshot AI -- coding-focused agentic model built on Kimi K2.6). They join the established 2026 lineup: Kimi K2.6 (Modified MIT, frontier MoE coding, SWE-Bench Pro 58.6), Qwen3.8-27B (best overall on consumer hardware, 61.7 SWE-bench Pro), GLM-5.3 (744B/40B active MoE, MIT, SWE-Bench Pro 58.4), and gpt-oss:20b (best small / 16 GB, ~o3-mini). Some newly listed models, such as mistral-large-4 and beam, are cloud-only (`:cloud` tag) and do not run locally. Pull commands: ollama run nemotron-3.5-lightning, ollama run nemotron-3.5-super, ollama run laguna-xs-2.1, ollama run kimi-k2.7-code:cloud, ollama run qwen3.8:27b.

Which Ollama models work best for Russian-language tasks?

Qwen3 / Qwen3.8 (`ollama run qwen3:8b` or `ollama run qwen3.8:27b`) and Mistral Small 3.1 (`ollama run mistral-small3.1`) both include Russian in their native multilingual training. This page does not rank the two for Russian, so test both on your own prompts. Neither is Russian-origin -- for regulatory data-residency requirements specific to Russia, check the model licence and your organization's compliance requirements separately, as this article does not cover Russian data-localization law.

Sources

A Note on Third-Party Facts

This article references third-party AI models, benchmarks, prices, and licenses. The AI landscape changes rapidly. Benchmark scores, license terms, model names, and API prices can shift between the time of writing and the time you read this. Before making deployment or compliance decisions based on this article, verify current figures on each provider’s official source: Hugging Face model cards for licenses and benchmarks, provider websites for API pricing, and EUR-Lex for current GDPR and EU AI Act text.

Run PromptQuorum with a local LLM, your own API keys, or both — you pick the backend.

Download the PromptQuorum Beta →

← Back to Local LLMs

MCP