Open-source AI in 2025: Model performance, enterprise adoption and investment

Editor’s Note

This article was adapted by EUVC from an original insight published by Synaptic.

Read the original article.

Choosing between open and proprietary AI is no longer simply a question of which model leads a benchmark.

For enterprises, the decision increasingly involves cost, control, security, customisation and deployment requirements. For investors, the opportunity is expanding beyond model developers into the infrastructure required to evaluate, monitor and operate AI in production.

Synaptic’s October 2025 market analysis captures this transition. It argues that proprietary models still lead the performance frontier, but open-source AI is becoming commercially relevant through a growing developer ecosystem, hybrid enterprise adoption and demand for production-ready infrastructure.

Throughout this article, “open-source AI” follows Synaptic’s market category. In practice, openness varies considerably: some models publish weights but not training data, development methods or unrestricted licences.

Synaptic’s 2024 analysis found that open models including Llama, Mistral and Qwen were narrowing the benchmark gap with proprietary systems.

Startup activity was also accelerating:

By 2025, the picture had become more complex. Proprietary models had extended their lead on demanding reasoning tests, while the commercial case for open models increasingly rested on cost, customisation and deployment flexibility.

Synaptic argues that this is encouraging a hybrid market rather than a winner-takes-all contest.

DeepSeek’s R1 family was one of the most important developments in Synaptic’s 2025 analysis.

The models used reinforcement learning to improve their ability to solve multi-step problems. Synaptic presents R1 as evidence that open-model developers can contribute to reasoning advances previously associated mainly with proprietary laboratories.

This did not erase the performance gap with leading closed models. It did, however, reinforce the role of open development in expanding access to model research and experimentation.

Compact, instruction-tuned models from developers including Meta and Alibaba also continued to attract downloads. Instruction tuning adapts a model to follow user requests more effectively, making smaller models more practical for specific applications.

Synaptic also highlights the risk of “open washing”: describing a model as open while withholding important components of its development or use.

A model may provide downloadable weights but restrict its commercial licence. It may disclose its architecture without publishing its training data, finetuning process or evaluation methodology.

The result is a spectrum rather than a binary distinction. When evaluating an open model, developers and investors need to examine:

This distinction matters because the benefits associated with open-source software cannot be assumed from an “open” label alone.

Enterprise adoption was becoming hybrid by 2025.

A McKinsey survey of more than 700 technology leaders and senior developers found that more than 50% of respondents used open-source technologies in each of the data, model and tooling layers of their AI stacks, often alongside proprietary products.

The trade-offs differ by workload. Proprietary frontier models may be preferred for demanding reasoning tasks, while smaller open models can offer greater control for specialised, cost-sensitive or on-premise applications.

An IBM-commissioned study of 2,413 IT decision-makers found that 51% of surveyed companies using open-source AI tools reported positive ROI, compared with 41% of those not using them. This was a reported correlation, not evidence that open-source adoption caused the difference.

The emerging pattern is therefore not the wholesale replacement of proprietary systems. It is workload-level selection across a mixed AI stack.

A related article on why individual AI adoption has outpaced enterprise-scale deployment provides additional context on the organisational barriers to putting these systems into production.

As AI adoption moves from experimentation to production, organisations need to know how their models behave after deployment.

Model monitoring tracks performance indicators such as accuracy, latency and drift. Observability goes further by helping teams investigate why an AI system produced a particular result, where failures occurred and how performance changes across data, prompts and model versions.

Synaptic identifies monitoring and observability as among the most active areas of the 2025 market. Earlier machine-learning monitoring companies such as Arize AI and Fiddler AI were expanding into broader platforms for evaluation, debugging and governance.

At the same time, companies including Baseten and LangChain were helping developers turn models into operational products through hosting, middleware and application frameworks.

This reflects a wider shift from developing models to making them reliable enough for production. For additional context on the physical compute layer supporting this market, see the related AI hardware market map.

Strategic acquisitions during 2025 showed large technology companies and infrastructure providers moving into adjacent parts of the AI stack.

Synaptic highlighted several examples:

These transactions suggest that evaluation, optimisation and data infrastructure were becoming strategically important alongside the models themselves.

Synaptic counted more than $3.5 billion across the open-source AI funding rounds covered in its 2025 analysis. Late-stage transactions accounted for much of the capital.

Three rounds illustrate the scale:

Earlier-stage funding was more distributed. Series A activity included companies working on AI infrastructure, data orchestration and computer vision, while seed capital flowed into prompt optimisation, model evaluation and monitoring.

Growth-stage infrastructure companies also attracted significant rounds. Synaptic highlighted fal and Arize AI as examples of investors backing platforms that had progressed beyond initial product validation.

This investment pattern supports Synaptic’s wider thesis: the commercial opportunity was moving from model development alone towards the systems required to deploy and manage models.

For a broader perspective on how investors divide the AI opportunity across knowledge, labour, machines and financial infrastructure, see this related AI investment framework.

Synaptic’s benchmark snapshot showed proprietary models leading on complex reasoning.

Gemini 2.5 Pro, Grok 4 and GPT-5 occupied the leading positions in the rankings it examined. Open models including GLM-4.5 and DeepSeek-V3.1 had improved but had not reached parity.

On Humanity’s Last Exam, a benchmark covering 2,500 questions across mathematics, logic, coding and other domains, Synaptic reported:

These figures represent a specific 2025 benchmark snapshot, not a permanent ranking. Model releases and evaluation methods change quickly, and benchmark performance does not capture every factor relevant to production deployment.

Usage data also showed a large gap between proprietary and open models.

In the OpenRouter snapshot used by Synaptic, the five most-used models accounted for approximately 12.9 trillion of 15.6 trillion tracked tokens, or 83% of usage. Grok Code Fast 1 alone represented 4.5 trillion tokens.

Open models nevertheless achieved meaningful adoption. DeepSeek V3.1 ranked sixth with 748 billion tokens, while the DeepSeek variants covered by the analysis collectively represented approximately 1.35 trillion tokens, or around 9% of tracked usage.

Hugging Face downloads told a similar story about developer interest. Meta’s Llama models and Alibaba’s Qwen family attracted substantial activity across different model sizes. Earlier model families including Falcon and Bloom had declined in relative popularity, which Synaptic interpreted as growing demand for efficient and flexible deployment.

GitHub activity also favoured a relatively small number of projects. Repositories associated with Anthropic, Supabase and LangChain recorded strong star growth during the second quarter of 2025, while foundational projects such as Hugging Face Transformers continued to attract attention.

Synaptic’s analysis points to four considerations for founders, operators and investors.

A model can trail the frontier while remaining useful for a specific application. Cost, latency, privacy, customisation and deployment location can matter as much as general-purpose reasoning performance.

Enterprises do not need to make a single permanent choice between open and proprietary AI. Different models can serve different workloads within the same organisation.

Monitoring, evaluation, data orchestration and model serving become more important when systems must operate reliably at scale. This creates opportunities beyond the model layer.

Licensing, access to weights, training transparency and infrastructure dependencies all affect how much control a user actually receives.

This also has implications for defensibility. A related framework on evaluating whether AI strengthens a software company or its underlying infrastructure providers offers further context for investors assessing where durable value may accrue.

Not across every task. In Synaptic’s 2025 snapshot, proprietary models led demanding reasoning benchmarks. Open models could still be competitive for specialised applications where cost, control or customisation mattered more than frontier performance.

Different workloads have different requirements. An organisation may use a proprietary model for complex reasoning and an open model for a specialised, on-premise or high-volume task.

AI observability is the ability to examine how an AI system behaves in production, including its outputs, failures, latency, data dependencies and changes between model or prompt versions.

Synaptic’s analysis points to model performance, deployment economics, developer adoption, licence terms, infrastructure requirements and the company’s ability to convert technical usage into sustainable enterprise demand.

Synaptic is the corporate author of the original article. The company provides private-markets data and intelligence, turning signals such as financial information, company materials, websites, news and alternative metrics into connected data for investors.

Learn more about Synaptic’s private-market research and data platform.

This article was adapted by EUVC from an original article published by Synaptic.

Read the original article.