Free large language models: the shift toward open-source AI

The capital is distributed across cloud providers, model labs, semiconductor companies, enterprise IT budgets and governments that would rather own an inference stack than rent one indefinitely.

The valuation story is less important than the cost structure. Open-weight models now cost, on average, 15.66% as much as closed alternatives. That does not make them free. GPUs, storage, networking, engineering and security still produce a sizeable burn rate. But it changes the bargaining position of every company buying model capacity.

The market is moving from “which chatbot is smartest?” to a harder question: who controls the model, the weights, the data path and the resulting economics?

The 13-week convergence: open models are catching up faster

The performance gap between open and proprietary large language models has not disappeared. It has become less useful as a strategic assumption.

In the first half of 2024, a leading open-source model needed an average of 27 weeks to match the performance of the best available closed model. By the first half of 2025, that lag had fallen to 13 weeks. The number matters because it measures more than benchmark progress. It measures how long a closed-model lead remains commercially defensible.

A six-month advantage can justify premium API pricing. A three-month advantage is more difficult to monetize, especially when the competing model can be downloaded, fine-tuned and deployed behind a company’s own firewall.

Open-weight models now reach roughly 90% of closed-model performance on leading evaluations such as GPQA, MMLU Pro and LiveCodeBench. That figure should not be mistaken for a universal quality score. Benchmarks differ in difficulty, contamination risk, task design and relevance to production workloads. A model that performs well on academic reasoning may still fail at structured extraction, tool use, multilingual support or long-running agent workflows.

But the direction is clear. Proprietary labs still hold advantages in several areas:

• Frontier training runs remain expensive and difficult to reproduce.
• Closed providers can tightly integrate models with search, tools, data stores and safety systems.
• Product telemetry gives them a rapid feedback loop from millions of users.
• Large customers may prefer contractual accountability over operating an internal model platform.

Open releases counter with different advantages:

• The weights can be inspected, adapted or fine-tuned.
• Inference can be moved between cloud, private infrastructure and local hardware.
• Organisations can control retention and data routing.
• Developers are not locked to one provider’s pricing, rate limits or product roadmap.

That trade-off is increasingly attractive for companies whose model usage is predictable. A legal-document pipeline, coding assistant or internal knowledge system does not need the most fashionable model every week. It needs stable quality at a known unit cost.

The open-model advantage is not that every model wins. It is that the buyer can change the rules after deployment.

This is where the phrase “free large language models” becomes economically misleading. The download may cost nothing. The operating model does not. Still, free access to model weights changes the cap table of the AI stack: fewer dollars flow automatically to the model vendor, while more spending moves toward infrastructure and engineering.

Economic realities: the price of independence

Open-weight models are estimated to cost 15.66% of the price of closed models on average. Put differently, closed models are approximately six times more expensive. Those multiples will vary sharply by workload, hardware utilisation, context length, output volume and service-level requirements. They remain useful because they expose the basic financial asymmetry.

A closed API usually bundles several costs into a per-token price:

1. Model training and ongoing research.

2. GPU depreciation and data-centre operations.

3. Network capacity and availability guarantees.

4. Safety, monitoring and abuse prevention.

5. Product development, support and margin.

A self-hosted or privately hosted open model removes the vendor’s margin and reduces exposure to per-token pricing. It also transfers risk to the customer. That risk appears on a different line of the income statement.

The buyer now has to fund:

• GPU procurement or reserved cloud capacity.
• Model serving and quantisation work.
• Observability, access controls and incident response.
• Patching and version management.
• Fine-tuning data preparation.
• Engineers who understand distributed inference.
• The opportunity cost of keeping an internal AI platform alive.

For a small team with low and volatile usage, a closed API may still be cheaper. Capital expenditure is not automatically efficient merely because it is private. Idle GPUs have poor liquidity. A rented service can be switched off; a hardware cluster cannot be liquidated quickly without taking a loss.

The calculation changes at scale. High-volume inference produces recurring API bills. If a company already operates data centres, has a large machine-learning team and faces strict data-residency requirements, open weights can improve gross margins while reducing dependency on one vendor.

The relevant comparison is therefore not “free versus paid”. It is variable cost versus fixed cost, vendor margin versus internal burn rate, and flexibility versus operational complexity.

The cost advantage also depends on the licence. “Open source” is often used as a convenient umbrella for models with very different rights. Some releases provide weights under permissive licences. Others use community licences with commercial restrictions. A downloadable foundation model is not necessarily an unrestricted commercial asset.

That distinction belongs in procurement meetings, not footnotes. A model may be technically downloadable while still imposing conditions on usage, redistribution, revenue thresholds or model derivatives. The legal and financial exposure can exceed the original inference bill.

The 2026 vanguard: Kimi K3, GLM-5.2 and DeepSeek-V4-Pro

The latest open-model releases illustrate how quickly the frontier is scaling.

Kimi K3, released in July 2026, is reported at 2.8 trillion parameters and offers a one-million-token context window. GLM-5.2, released in June under the MIT licence, is a 744-billion-parameter mixture-of-experts model. DeepSeek-V4-Pro, released in April, is a 1.6-trillion-parameter MoE model. DeepSeek also released a V4-Flash variant during the same period.

Parameter counts are useful shorthand but weak investment analysis. They indicate the size of the model architecture, not the total training bill, active parameters per token, serving cost, data quality or production reliability. A sparse mixture-of-experts system may contain a very large total parameter count while activating only part of the network for each request.

The more meaningful questions are operational:

• How many parameters are active during inference?
• What hardware does the model require at acceptable latency?
• Does quantisation preserve the quality needed by the target workload?
• Is the licence compatible with commercial deployment?
• Are the weights, tokenizer, evaluation code and serving tools actually available?
• Can the model handle the required languages, context length and tool calls?
• How often will the publisher release fixes or updated checkpoints?

Kimi K3’s one-million-token context window is a notable capability, but context capacity is not the same as context utility. A model can accept a large prompt and still lose accuracy deep inside it, retrieve the wrong passage or become too expensive to serve. Long context also creates a pricing problem: the theoretical ability to process more information can become a practical incentive to send too much information.

GLM-5.2’s MIT licence is more commercially legible than a licence with extensive field-of-use restrictions. That matters to enterprise buyers because a clean licence reduces friction in the legal review. It does not eliminate technical or compliance obligations, but it improves the model’s position on a procurement spreadsheet.

DeepSeek’s releases represent another pressure point. The company’s earlier progress helped demonstrate that competitive reasoning systems could emerge outside the established Western lab structure. Its newer models push the same argument into scale. The market no longer treats open releases as merely cheaper substitutes for older proprietary models. It increasingly treats them as a parallel supply chain.

That is a financial threat to closed providers. If a model’s quality is adequate and its weights are available, the provider must compete on more than raw intelligence. It needs stronger tooling, better uptime, easier compliance, superior distribution or a lower total cost of ownership.

A benchmark score alone cannot settle that contest. Model evaluation has become a due-diligence problem. The industry needs reproducible tests, disclosed hardware assumptions and workload-specific measurements. Otherwise, a release announcement is just another asset-marketing document.

Some claims in the AI market now deserve the same treatment as guides to credible rumor sources: separate the verifiable event from the attractive story built around it. A published benchmark is evidence. A claim that the model “matches the frontier” is an interpretation. The distinction is where the money is won or lost.

Sovereignty and security: why on-premises deployment is gaining ground

On-premises LLM deployments accounted for 51.85% of the market in 2025. The figure reflects more than a preference for private servers. It reflects a concern about control.

Enterprises are under pressure to keep sensitive data within defined jurisdictions, limit third-party access and control the latency of business-critical systems. A closed API can meet those requirements in some configurations. It cannot meet them by default for every organisation, every regulator and every data class.

Open models give the customer more options:

• Run inference in an internal data centre.
• Use a private cloud account with controlled network routes.
• Place smaller models on edge or departmental hardware.
• Fine-tune on proprietary data without sending the training set to a model provider.
• Retain logs and prompts under internal governance.
• Freeze a model version instead of accepting silent changes to a hosted system.

This is not a promise of perfect security. Self-hosting creates a larger attack surface if the organisation lacks the staff to patch infrastructure, protect model endpoints and monitor misuse. Model weights can be stolen. Fine-tuning data can leak. An internal deployment can fail just as expensively as an external one, only without a vendor’s support team to blame.

Sovereignty is also a question of bargaining power. A company that depends on one proprietary model provider may face abrupt price changes, usage limits or an API redesign. A company that can operate several open models has more negotiating leverage. It can shift workloads according to cost, latency and performance.

The strongest architecture will often be hybrid rather than ideological. A private model can handle sensitive documents and predictable internal tasks. A closed frontier model can handle specialised reasoning, high-end multimodal work or temporary demand spikes. The cap table of providers becomes a portfolio instead of a single holding.

The risk is fragmentation. Running five model families may reduce vendor concentration while increasing internal complexity. Each model brings different tokenisation, prompting behaviour, context limits, evaluation quirks and licence conditions. Switching is possible. It is not free.

Strategic integration: how enterprises are adopting open AI

Around 63% of organisations already use open-source AI somewhere in their technology stack. Broader estimates put enterprise usage of at least one open-source AI model at 75%.

These numbers do not mean three-quarters of companies are replacing closed providers with self-hosted systems. They likely include open-source libraries, machine-learning frameworks, embedding models, development tools and open-weight models used in limited production settings. The distinction matters. Adoption is broad; deep operational dependence is not necessarily broad.

Still, the strategic shift is visible. Companies are adopting open models in layers.

1. Development and experimentation

Developers use free AI models for prototyping, code assistance, classification and retrieval pipelines. The low entry cost accelerates testing. Teams can compare architectures without negotiating a new enterprise contract for every experiment.

This stage creates technical familiarity. It also creates hidden liabilities. A prototype that depends on an unreviewed licence or a model with uncertain provenance can become an expensive rewrite when it reaches production.

2. Narrow production workloads

Open models are most financially compelling when the task is stable and measurable. Document classification, summarisation, extraction and internal search often fit that profile. The model does not need to win every benchmark. It needs to meet a defined accuracy threshold at an acceptable cost.

In these applications, a smaller model with predictable latency can outperform a larger proprietary model on total economics. The comparison should include human review rates, error correction, infrastructure utilisation and failure handling. Token price alone is a poor proxy for business value.

3. Fine-tuning and domain adaptation

Open weights allow companies to adapt a model to internal terminology, formats and workflows. Fine-tuning can improve consistency, but it also consumes data, compute and specialist labour. The resulting model may become an internal asset, yet its value depends on how often the business process changes.

Fine-tuning is not a magic multiplier. Poor data produces a more confidently wrong system. Narrow training can reduce general capabilities. Version updates may force the organisation to repeat the work, creating a maintenance burden that does not appear in the initial deployment budget.

4. Private inference at scale

Large organisations with substantial request volumes can bring inference closer to the data and optimise serving for their workload. Quantisation, batching and specialised hardware can reduce cost. The benefit increases when GPUs remain highly utilised.

This is where the 15.66% cost figure becomes most relevant. A large buyer can potentially capture the spread between an open model’s infrastructure cost and a closed provider’s retail price. But the spread is not pure margin. It must absorb staff, power, hardware refresh cycles and downtime.

5. Portfolio deployment

The mature strategy is rarely one model for everything. Enterprises are beginning to treat models as a portfolio of assets with different cost and capability profiles. A lightweight open model may handle routine tasks. A larger model can process complex cases. A proprietary system may remain in the stack for high-stakes reasoning or multimodal workloads.

That architecture reduces dependence on any one vendor but makes evaluation continuous. The organisation needs internal benchmarks tied to actual business outcomes. Generic leaderboard performance is useful for screening. It is not a substitute for production telemetry.

What the market is really pricing

Forecasts put open-source LLM market growth at $70.23 billion between 2026 and 2030, with a compound annual growth rate of 34.1%. The projection is substantial. It is also a forecast, not liquidity in the bank.

The central commercial question is who captures that growth. Model publishers may distribute weights for free while monetising hosted inference, enterprise support, fine-tuning, hardware partnerships or cloud consumption. Cloud providers can benefit even when model licence fees collapse, because every self-hosted model still needs compute somewhere. Chip companies benefit from the same dynamic. Systems integrators may capture the deployment work.

This is why “free” can be a powerful distribution strategy. Giving away weights can shift the revenue pool rather than eliminate it. The model becomes the entry point for a larger infrastructure and services business.

The pressure on proprietary labs will be uneven. The best-funded providers can sustain high training costs if their products retain a quality, distribution or ecosystem advantage. Smaller closed-model companies face a harsher multiple. If open releases narrow the capability gap while undercutting price, a modest performance lead may not support a premium valuation.

Investors will look for evidence in three places:

• Gross-margin durability: Can the company maintain pricing when open alternatives improve every quarter?
• Retention and switching costs: Do customers stay because of workflow integration, or merely because migration is inconvenient?
• Infrastructure capture: Does the company control a profitable layer of the stack, or does it pay others for compute while competing on model price?

The open-source model boom therefore does not imply that model labs will become unprofitable by default. It implies that the market is becoming less tolerant of businesses whose only asset is access to a model that competitors can approximate within 13 weeks.

The limits of the open-model thesis

There are three easy mistakes in the current narrative.

The first is treating every open-weight release as a pure public good. It is not. Compute remains expensive. Inference remains an operating cost. Large models can require specialised clusters, and even smaller systems need reliable serving, security and maintenance.

The second is assuming that open models have already replaced closed ones. They have not. Closed providers still control significant demand, especially where customers value managed infrastructure, multimodal features, compliance tooling and contractual service guarantees.

The third is confusing technical access with strategic independence. An enterprise may download a model and still depend on one cloud provider, one GPU supplier, one inference framework and a small group of engineers. The vendor lock-in has moved down the stack. It has not vanished.

There is also a governance risk. Open distribution can accelerate legitimate research and commercial deployment, but it can also make harmful capabilities easier to reproduce. Enterprises need access controls, usage monitoring and model-risk processes whether the weights came from a public repository or a paid API.

The winning organisations will not be the ones that repeat “open” most loudly. They will be the ones that can show a lower total cost, acceptable reliability and a credible path to maintain the system after the launch attention disappears.

A cheaper model is not automatically a better asset

Free large language models are changing the economics of AI procurement. They reduce the premium that proprietary providers can charge for basic language capability. They give developers more control over data and deployment. They compress the time available for closed-model leaders to turn a performance advantage into durable revenue.

But the open-source shift is not a victory parade. It is a transfer of risk.

The customer saves on API margins and assumes more responsibility for infrastructure, licensing, security and model operations. Model publishers gain adoption but must find monetisation beyond the weights. Cloud and hardware companies may capture the spending that model labs lose. Investors will have to decide which layer has pricing power when the model itself becomes abundant.

The sober conclusion is straightforward. Open models are becoming strategically important because they make AI capacity more contestable. They do not make AI free. The bill has simply moved from the model endpoint to the rest of the stack.