On 9 July 2026, Ollama announced the close of a $65 million Series B funding round led by Theory Ventures, with participation from Benchmark, 8VC, Y Combinator, Pace Capital, 49 Palms, and GTMFund. Total funding now stands at $88 million, according to the company. Founded in 2023, Ollama built the most widely used tool for running large language models locally, and this round signals that the market it helped create has become too large for institutional investors to ignore.
What Ollama is, and why the numbers matter
Ollama is an open-source platform that lets anyone run large language models on their own hardware. A single terminal command downloads and launches models like Qwen3, Llama 4, Gemma 4, or Mistral Large 3, no cloud API, no data transfer, no per-user subscription. The platform handles model quantisation, prompt formatting, and a local OpenAI-compatible REST API, making it trivially easy to drop local inference into existing applications.
According to the company, 8.9 million developers use the platform every month. Usage doubled since January 2026 alone, and new installs are adding roughly one million per week as reported at the time of the funding announcement. Over 67,000 integrations have been built on top of Ollama, IDE plugins, RAG pipelines, automation frameworks, enterprise applications. Perhaps most striking: 85% of Fortune 500 companies are reportedly using Ollama in some capacity, including organisations in healthcare, government, and financial services that are not typically known for adopting early-stage open-source tooling.
The inflection point arrived around January 2026, when open-weight models crossed the quality threshold required for reliable agentic workflows, the kind where the model calls tools, makes sequential decisions, and drives multi-step processes without human intervention in every loop.
ollama launch: local models for your coding assistant
Alongside the funding news, Ollama shipped a feature that directly changes the economics of AI-assisted development for privacy-sensitive teams: the ollama launch command. It configures coding assistants, Claude Code, OpenCode, and Codex, to communicate via a local OpenAI-compatible API (localhost:11434) using a model you choose. No environment variables, no configuration files, no outbound connections.
Requirements: Ollama v0.15 or newer. Recommended context length: at least 64,000 tokens, since coding assistants perform significantly better with large code contexts. The strongest local options for coding tasks, according to community benchmarks, are Gemma 4 26B (reported at 77.1% on LiveCodeBench v6) and Qwen3.5 27B (reported at 80.7%). Both run on a Mac Studio M3 Ultra or a single high-VRAM GPU without quantisation that visibly degrades quality.
For development teams working with proprietary source code, client codebases, or regulated data, this matters: with ollama launch, not a single line of code leaves your network during a coding session.
July 2026 technical updates
The same release cycle brought several under-the-hood improvements that affect real-world hardware setups:
- Flash attention for older NVIDIA GPUs (compute capability 6.x): GPUs from the GTX 1000 generation now benefit from accelerated attention computation, with meaningfully shorter generation times.
- Gemma 4 roughly 90% faster on Apple Silicon: Multi-token prediction is automatically enabled on compatible Macs. It requires no configuration and does not change the model's output.
- iGPU vision offload: Integrated graphics can now load vision models with padding to fit available memory, useful for compact, GPU-less setups.
- Web search and web fetch plugin for OpenClaw: Local models can retrieve and process live web content in-context without the model itself needing internet access.
These additions position Ollama less as a bare runtime and more as a full local AI platform with a growing plugin layer, a pattern that mirrors the evolution of other category-defining open-source infrastructure projects.
What the funding round signals for the local AI market
A $65M Series B from investors with long track records in open-source infrastructure is not a bet on a niche. It is a bet that private, on-premise AI inference is a durable infrastructure layer, one that will grow alongside, not instead of, cloud AI.
Ollama's strategy reflects this. The company runs a hybrid model: complete data sovereignty for users who want fully local operation, and a scalable cloud option for teams with variable capacity needs. Cloud token usage reportedly grew over 200% month-on-month at the time of the announcement, yet the local open-source product remains the stated core of the strategy.
For European SMBs operating under GDPR or preparing for the EU AI Act's deployer obligations, the local path has a distinct compliance advantage: data that never leaves your infrastructure is data that never triggers cross-border transfer restrictions, external processor agreements, or input-logging concerns. Our overview of local AI for European businesses covers what this means in practice across different industries.
The funding round also carries a longevity signal: Ollama is now capitalised well enough to outlast the typical open-source project funding gap. Companies building internal workflows on Ollama today are less exposed to the platform abandonment risk that has historically made enterprise architects cautious about open-source AI tooling.
Practical guidance for SMBs considering local AI
The hardware landscape in mid-2026 makes entry more accessible than it was a year ago. The following configurations have proven reliable in production contexts, based on community-reported benchmarks:
- Mac Mini M4 Pro (48 GB): Entry point for 7B-14B models, near-silent, under 30 W at load. Good for document processing, internal chatbots, and summarisation.
- Mac Studio M3 Ultra (192 GB): 70B models at 25-40 tokens/second as reported by practitioners. Suitable for team-shared inference, RAG over large document sets, and agentic workflows.
- RTX 4090 (24 GB VRAM): Windows-compatible alternative under €2,000; Qwen3 8B and Gemma 4 12B run at 30-60 tokens/second.
For sectors with strict data handling requirements, legal, healthcare, accounting, a local deployment is often not one option among several; it is the only configuration that keeps sensitive client or patient data unambiguously within your infrastructure. Our data sovereignty page explains how this maps onto GDPR Article 28 processor obligations and the EU AI Act's transparency requirements.
The bigger picture: infrastructure maturity
67,000 integrations, 8.9 million active developers, and $65M in fresh institutional capital. These figures no longer describe a hobbyist tool or a niche privacy project. They describe infrastructure that has reached the scale at which it becomes a credible long-term dependency for enterprises.
The ollama launch feature captures something important about where local AI is heading: it is no longer a privacy compromise that sacrifices convenience. With models hitting over 80% on coding benchmarks and running locally on hardware that fits on a desk, the productivity case is closing in on the cloud-based alternative.
Organisations that are still evaluating whether local LLMs are worth exploring are past the point of early-mover advantage, but there is still meaningful compounding benefit in starting now.
Want to assess what a local AI setup would look like for your team? Our pilot project programme gets you to a working proof of concept in two to four weeks, with defined success criteria and a clear path to production.