Ollama July 2026 Update: v0.32.0 + Best Models by Use Case

Current Ollama version: v0.32.0 (released July 11, 2026). This is the latest stable release, available via ollama.com/download. Update with `curl https://ollama.ai/install.sh | sh` (macOS: `brew upgrade ollama`), then confirm with `ollama --version`.

What changed in the v0.31–v0.32 series (late June–July 2026): v0.30.11 (June 25) added auto-install for Claude Code and opencode and fixed Windows hybrid-GPU classification. v0.30.12 (June 29) fixed tool-call detection in JSON strings. v0.31.1 (June 30) shipped a new MLX batch-matmul kernel that makes Gemma 4 generate tokens nearly 90% faster on Apple Silicon via multi-token prediction. v0.31.2 (July 6) enabled flash attention on older NVIDIA GPUs and fixed structured output for thinking models. v0.32.0 (July 11) introduced an interactive "Chat, Code & Work" agent experience and renamed the Codex App integration to ChatGPT. Full notes: github.com/ollama/ollama/releases.

Newest models added since the last refresh (late June–July 2026):

Popularity on Ollama is measured by download counts visible on each model's library page. As of July 2026, the top downloaded models are still dominated by Meta's Llama family -- Llama 3.2 3B is the most pulled model overall, largely due to its use as a first-install test model. However, Llama 4 Scout has climbed rapidly since its April 2026 release.

Qwen3 is the fastest-growing model family in the Ollama library, with Qwen3 and the new Qwen 3.6 dense variant quickly displacing Qwen2.5. DeepSeek-R1 saw major spikes following its release and remains highly downloaded for reasoning tasks. Dolphin 3.0 (Cognitive Computations, built on Llama 3.1) remains the most-pulled uncensored general-purpose model for users who want an unfiltered local assistant.

Meta released Llama 4 in April 2026 with Scout (17B active, 109B total, MoE) and Maverick (17B active, 400B total) variants. Llama 4 Scout is now stable in the Ollama library (`ollama pull llama4:scout`). The Llama 4 family uses Mixture-of-Experts (MoE) architecture — Scout needs ~55 GB VRAM at Q4 (it fits 24 GB only at 1.78-bit, ~20 tok/s).

Scout's standout features are its 10M-token context window and multimodal input, not consumer-hardware fit. For lightweight setups (8 GB RAM), Llama 3.2 3B remains the easiest first model.

Kimi K2.6 (Moonshot AI, Modified MIT license, 32B active / 1T total MoE) scored SWE-Bench Pro 58.6, tying GPT-5.5. Laguna XS 2.1 (Poolside) is the newest agentic-coding challenger, purpose-built for long-horizon local coding sessions. Qwen 3.6 27B achieved 77.2% SWE-bench as the best overall model on consumer hardware (fits 24 GB at Q4). OpenAI's open-weight gpt-oss:20b (21B total / 3.6B active MoE) runs in 16 GB at ~o3-mini level with adjustable reasoning.

The Ollama library holds several hundred curated models, with Laguna XS 2.1 and Kimi K2.7 Code the newest additions since the June refresh.

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.

These are the newest models in the Ollama library as of July 2026, newest first. Confirm availability with `ollama pull <model>` before building workflows — new models appear at ollama.com/library within days of release.

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 scores 52% on MATH (competition math) vs 28% for Mistral Small at the same size. It is slower than standard models (more tokens per response) but significantly more accurate on tasks where reasoning matters.

As of July 2026, 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.

Download counts still favor Llama 3.x due to tutorial prevalence. For new projects in July 2026, prefer Qwen 3.6 27B (best overall on consumer hardware), Kimi K2.6 or Laguna XS 2.1 (coding), gpt-oss:20b, and qwen3:30b.

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.

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, Mistral Small 3.1, Qwen3 7B, Qwen 3.6 27B, Devstral Small 24B, Gemma 2 2B. Meta Llama Community Licence (commercial use restricted above 700M monthly active users): Llama 3.3 8B, Llama 3.2 3B, Llama 3.2 Vision 11B. 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. For EU enterprises in regulated sectors, Mistral models (France, Apache 2.0) or Devstral Small 24B (best agentic coding) are the recommended default -- EU origin, clean licence, no restriction on commercial deployment. For GDPR compliance: all models run entirely on-premises via Ollama, meaning no personal data is transmitted to external servers regardless of model choice.

Japan (METI). For Japanese enterprise Ollama deployments, Qwen3 / Qwen 3.6 is the recommended model family -- native Japanese tokenization processes Japanese text 30-40% more token-efficiently than Llama or Mistral, directly reducing inference time and KV cache requirements. For Japanese coding workflows: Qwen 3.6 27B (77.2% SWE-bench) handles Japanese code comments natively and is the top dense coding model in 2026. METI AI governance documentation requires noting the exact model version. Use `ollama show <model>` to get the full model specification including parameter count, quantization level, and context length for compliance records.

China. Under China's CAC Generative AI Measures (2023), organizations providing AI services to end users must register the models used. Qwen3 / Qwen 3.6 (Alibaba, Apache 2.0) is the recommended choice for Chinese enterprise Ollama deployments -- Chinese model origin, Apache 2.0 licence, best performance on Chinese-language tasks, and top benchmarks. Kimi K2.6 (Moonshot AI, Modified MIT license, 32B active/1T total MoE) is also available as a top-tier coding option with Chinese origin. Pull commands: `ollama run qwen3.6:27b` for best quality, `ollama run qwen3:7b` for speed. DeepSeek-R1 (DeepSeek, MIT licence) is appropriate for reasoning tasks. For data processed locally via Ollama, China's PIPL cross-border data transfer requirements do not apply -- inference stays on-premises.