An AI that runs locally could give small businesses more bargaining power. That promise depends on whether they can keep using, repairing and replacing it.
On October 27, Liquid AI released LFM3 (https://www.liquid.ai/blog/lfm3), a family of downloadable AI models designed for specialised work on devices. This establishes a new option for developers, not proof that every inexpensive phone can deliver a dependable service. The consequential possibility is economic: people may gain an alternative when a distant provider raises prices, changes terms or withdraws access. Small AI should be judged by how much practical independence it gives its users.
“Open weights” means developers can download the model itself and run or adapt it, subject to its licence. Liquid's text-model documentation (https://huggingface.co/LiquidAI/LFM3-1.7B) lists Hindi among its languages, while warning against complex, multi-step reasoning. Its licence (https://huggingface.co/LiquidAI/LFM3-1.7B/raw/main/LICENSE) also restricts commercial use above a revenue threshold. Availability, suitability and permission are separate questions.
Imagine a small Indian distributor using a locally installed application to turn photographed invoices into draft entries. If it works reliably, a clerk could continue preparing records during a connection outage, check the amounts and decide what gets accepted. Keeping that task operational without a live subscription would have value even if the firm usually preferred a cloud service. This is a possible application, not a reported LFM3 deployment.
The bargaining mechanism is familiar. A customer who can leave has more influence over the terms on which they stay. But downloading a model creates little leverage if business records remain trapped in a supplier's format, installation requires scarce expertise, or replacing the application breaks daily work. Independence has to include the surrounding service.
The World Bank's 2025 AI foundations report (https://www.worldbank.org/en/publication/dptr2025-ai-foundations) identifies opportunities in small AI while emphasising infrastructure, data and skills. That combination matters. Computation may move nearer the user, but someone still pays for equipment, electricity, testing, updates and mistakes. A subscription bill can disappear while a maintenance burden quietly arrives.
For workers, the result depends on how that burden is allocated. A clerk might gain time for customers and authority to correct bad outputs. Alternatively, an employer might raise throughput targets and leave the clerk personally responsible for errors. Local implementation and support could create work for nearby businesses, but that is a potential market, not evidence of net job creation. Training and correction time belong in the budget.
India also exposes the limits of equating local computing with personal control. In the NSO's January–March 2025 telecom survey (https://www.pib.gov.in/Pressreleaseshare.aspx?PRID=2132330&lang=2®=3), 48.4% of rural women aged 15 and above owned a mobile phone, compared with 80.7% of rural men. These figures concern all mobile phones under the survey's ownership definition, not devices capable of running a new AI model. They describe an access gap that a smaller download cannot close.
A shared household phone may offer convenience without privacy. An employer's device may keep information away from a cloud provider while giving the employer greater power to monitor staff. An interface that fails with a person's language or disability can make nominal access unusable. Where computation happens matters; who controls the device and application matters just as much.
The strongest objection is that most people do not want to become their own IT department. Cloud providers can pool maintenance, security and computing resources, and larger models may perform difficult tasks better. Requiring local operation everywhere could make services more expensive and less reliable. A credible alternative should improve the choice to use cloud services, without turning self-sufficiency into an obligation.
Three futures follow. In a constructive one, cooperatives and small firms pool testing and support, keep portable records and buy tools they can replace. In an uneven one, technically capable firms gain bargaining power while everyone else remains dependent. In a harmful one, inexpensive local AI expands monitoring inside workplaces and homes while closed applications keep users locked in. None is predetermined by model size.
India's competition commission recommended improving access to AI infrastructure and capabilities (https://www.pib.gov.in/PressReleaseIframePage.aspx?PRID=2175607&lang=2®=48) in its 2025 study release. Practical implementation could include shared testing facilities and procurement demonstrations in which a buyer actually exports records and changes suppliers. Device access and paid support deserve attention alongside the model download.
Questions for founders and investors
Can customers keep essential work running if your service disappears? Can another provider use their records? What does a completed, corrected task cost on the devices they actually own? Who pays for maintenance, and who retains the authority to reject an answer?
Build and finance those capabilities. The most valuable small AI may be the one that lets a customer say no—and still open for business tomorrow.
Sources
Liquid AI release, October 27, 2026 (https://www.liquid.ai/blog/lfm3) · Model documentation (https://huggingface.co/LiquidAI/LFM3-1.7B) · Model licence (https://huggingface.co/LiquidAI/LFM3-1.7B/raw/main/LICENSE) · World Bank, 2025 (https://www.worldbank.org/en/publication/dptr2025-ai-foundations) · NSO telecom survey, 2025 (https://www.pib.gov.in/Pressreleaseshare.aspx?PRID=2132330&lang=2®=3) · CCI study release, 2025 (https://www.pib.gov.in/PressReleaseIframePage.aspx?PRID=2175607&lang=2®=48)
On October 27, Liquid AI released LFM3 (https://www.liquid.ai/blog/lfm3), a family of downloadable AI models designed for specialised work on devices. This establishes a new option for developers, not proof that every inexpensive phone can deliver a dependable service. The consequential possibility is economic: people may gain an alternative when a distant provider raises prices, changes terms or withdraws access. Small AI should be judged by how much practical independence it gives its users.
“Open weights” means developers can download the model itself and run or adapt it, subject to its licence. Liquid's text-model documentation (https://huggingface.co/LiquidAI/LFM3-1.7B) lists Hindi among its languages, while warning against complex, multi-step reasoning. Its licence (https://huggingface.co/LiquidAI/LFM3-1.7B/raw/main/LICENSE) also restricts commercial use above a revenue threshold. Availability, suitability and permission are separate questions.
Imagine a small Indian distributor using a locally installed application to turn photographed invoices into draft entries. If it works reliably, a clerk could continue preparing records during a connection outage, check the amounts and decide what gets accepted. Keeping that task operational without a live subscription would have value even if the firm usually preferred a cloud service. This is a possible application, not a reported LFM3 deployment.
The bargaining mechanism is familiar. A customer who can leave has more influence over the terms on which they stay. But downloading a model creates little leverage if business records remain trapped in a supplier's format, installation requires scarce expertise, or replacing the application breaks daily work. Independence has to include the surrounding service.
The World Bank's 2025 AI foundations report (https://www.worldbank.org/en/publication/dptr2025-ai-foundations) identifies opportunities in small AI while emphasising infrastructure, data and skills. That combination matters. Computation may move nearer the user, but someone still pays for equipment, electricity, testing, updates and mistakes. A subscription bill can disappear while a maintenance burden quietly arrives.
For workers, the result depends on how that burden is allocated. A clerk might gain time for customers and authority to correct bad outputs. Alternatively, an employer might raise throughput targets and leave the clerk personally responsible for errors. Local implementation and support could create work for nearby businesses, but that is a potential market, not evidence of net job creation. Training and correction time belong in the budget.
India also exposes the limits of equating local computing with personal control. In the NSO's January–March 2025 telecom survey (https://www.pib.gov.in/Pressreleaseshare.aspx?PRID=2132330&lang=2®=3), 48.4% of rural women aged 15 and above owned a mobile phone, compared with 80.7% of rural men. These figures concern all mobile phones under the survey's ownership definition, not devices capable of running a new AI model. They describe an access gap that a smaller download cannot close.
A shared household phone may offer convenience without privacy. An employer's device may keep information away from a cloud provider while giving the employer greater power to monitor staff. An interface that fails with a person's language or disability can make nominal access unusable. Where computation happens matters; who controls the device and application matters just as much.
The strongest objection is that most people do not want to become their own IT department. Cloud providers can pool maintenance, security and computing resources, and larger models may perform difficult tasks better. Requiring local operation everywhere could make services more expensive and less reliable. A credible alternative should improve the choice to use cloud services, without turning self-sufficiency into an obligation.
Three futures follow. In a constructive one, cooperatives and small firms pool testing and support, keep portable records and buy tools they can replace. In an uneven one, technically capable firms gain bargaining power while everyone else remains dependent. In a harmful one, inexpensive local AI expands monitoring inside workplaces and homes while closed applications keep users locked in. None is predetermined by model size.
India's competition commission recommended improving access to AI infrastructure and capabilities (https://www.pib.gov.in/PressReleaseIframePage.aspx?PRID=2175607&lang=2®=48) in its 2025 study release. Practical implementation could include shared testing facilities and procurement demonstrations in which a buyer actually exports records and changes suppliers. Device access and paid support deserve attention alongside the model download.
Questions for founders and investors
Can customers keep essential work running if your service disappears? Can another provider use their records? What does a completed, corrected task cost on the devices they actually own? Who pays for maintenance, and who retains the authority to reject an answer?
Build and finance those capabilities. The most valuable small AI may be the one that lets a customer say no—and still open for business tomorrow.
Sources
Liquid AI release, October 27, 2026 (https://www.liquid.ai/blog/lfm3) · Model documentation (https://huggingface.co/LiquidAI/LFM3-1.7B) · Model licence (https://huggingface.co/LiquidAI/LFM3-1.7B/raw/main/LICENSE) · World Bank, 2025 (https://www.worldbank.org/en/publication/dptr2025-ai-foundations) · NSO telecom survey, 2025 (https://www.pib.gov.in/Pressreleaseshare.aspx?PRID=2132330&lang=2®=3) · CCI study release, 2025 (https://www.pib.gov.in/PressReleaseIframePage.aspx?PRID=2175607&lang=2®=48)