A bank in Ho Chi Minh City points a general-purpose LLM at a stack of Vietnamese KYC documents. A hospital in Jakarta asks an off-the-shelf model to summarize patient charts. An insurer in Bangkok wires GPT into its claims workflow because it’s the fastest way to ship a demo. All three run the same playbook: take the biggest, most capable model on the market, prompt it well, and hope accuracy follows.
For a growing number of Southeast Asian enterprises, it doesn’t. And the gap between “the model passed our demo” and “the model works in production” has become the single biggest tax on generative AI ROI in the region.
The one-size-fits-all model is running out of road
The scale of the problem is no longer anecdotal. Industry trackers now put enterprise AI initiative failure rates as high as 85%, with a widely cited estimate suggesting that around 95% of generative AI pilots deliver no measurable P&L return once they leave the sandbox (Syntes.ai, 2026). S&P Global has reported that 42% of firms abandoned their primary AI initiative between 2025 and 2026 specifically because they could not prove a path to ROI.
The common thread in most post-mortems isn’t that the underlying model is weak — frontier LLMs are extraordinarily capable in general. It’s that a generic, globally-trained model is being asked to make deterministic, high-stakes decisions inside a specific business context it was never trained on: local language, local regulation, local document formats, internal taxonomy, and edge cases that only show up in a company’s own historical data.
Gartner has been tracking the shift away from generic deployment for exactly this reason, forecasting that more than half of enterprise generative AI deployments will be domain-specific by 2027 — up from roughly 1% in 2024. That is not a niche trend; it is a re-platforming of how enterprises intend to run AI in production.
Where the accuracy gap actually comes from
For Southeast Asian enterprises, the gap has two layers that compound each other.
The language and regional layer. Most frontier LLMs are trained overwhelmingly on English and Chinese data, with Southeast Asian languages treated as an afterthought. Independent benchmarking on MMLU-ProX has found accuracy gaps of up to 24.3% between high-resource and low-resource languages on the exact same questions — and Vietnamese, Bahasa, Thai, and Khmer all sit on the losing side of that gap. Region-specific models close it: on the M3Exam world-knowledge benchmark, SeaLLMs-v3-7B scored 64.9% in Vietnamese versus 62.4% for the general-purpose Qwen2-7B-Instruct and just 51.3% for Sailor-7B. On Vietnamese math reasoning (MGSM), SeaLLMs-v3-7B reached 71.2% against Meta-Llama-3-8B-Instruct’s 46.8% (SeaLLMs 3, arXiv 2024). Same task, same difficulty, wildly different reliability — purely because of how the base model was trained.
The domain layer. Even a model that handles Vietnamese well doesn’t automatically understand a bank’s KYC checklist, an insurer’s claims taxonomy, or a hospital’s discharge-note format. Generic models pattern-match on the closest thing they’ve seen in public data, which for anything document-heavy and locally regulated is often not close enough. This is the layer where hallucination and misclassification quietly creep into “high-confidence” outputs — the kind that look fine in a demo and fail in an audit.
Stack a regional gap on top of a domain gap, and it becomes obvious why so many SEA production deployments underperform their pilots: the model was tested on English- or Chinese-flavored public benchmarks and then handed a workload it was never actually trained for.
Why a smaller, fine-tuned model tends to win
Here is the counterintuitive part: closing the gap usually doesn’t require a bigger model. It requires a more specific one.
Parameter-efficient fine-tuning methods like LoRA and QLoRA can cut the compute and data required to specialize a model by more than 90% compared to full retraining, letting a team adapt a 7B–14B model on a handful of GPUs rather than a data center. In head-to-head internal evaluations, fine-tuned small and mid-size models have been shown to outperform larger general-purpose LLMs specifically on structured, repeatable, domain-defined tasks — the exact shape of work that dominates KYC review, claims triage, and clinical documentation (Forbes Technology Council, 2026). In financial services specifically, fine-tuned smaller models have been used to hit sub-100ms response times for fraud detection at near-zero marginal inference cost — a latency and cost profile that calling a hosted frontier model API cannot match.
That’s the pattern SEA enterprises are converging on: instead of routing every Vietnamese KYC document, every Bahasa customer ticket, or every Thai claims form through an expensive general-purpose API, fine-tune a smaller, regionally-aware base model on your own historical data. The result is typically the same shape every time, accuracy goes up because the model has actually seen your document types before, latency goes down because the model is smaller and can be deployed close to your data, and cost per inference drops because you’re no longer paying frontier-model API rates for a narrowly-scoped task.
What this looks like in practice on GreenNode AI Platform
AI Platform now splits model access into exactly the two lanes this article has been building toward:
- Third-party models, called through an API. Via AI Gateway, teams can reach hosted models like Kimi, MiniMax, Qwen 3.x, and GLM without integrating each provider separately — fast to start, fine for prototyping and lower-stakes tasks. But you’re renting someone else’s weights: no fine-tuning on your own data, and every request still pays that provider’s API economics.
- Self-hosted models, deployed on your own GPU allocation. These run on AI Platform’s own infrastructure (currently including Gemma), which is what makes fine-tuning possible in the first place. And self-hosted isn’t limited to a fixed dropdown: Model Registry’s vLLM and custom-container import paths let a team fine-tune an open-weight, regionally-strong base — SeaLLMs, Qwen2.5, or another open checkpoint — in a Notebook, then bring that specific checkpoint in for deployment rather than being boxed into whatever ships in the quick-deploy catalog.
The rest of the pipeline is the same regardless of which self-hosted base a team starts from:
- Fine-tune without standing up infrastructure. Notebook instances give AI engineers a GPU-backed Jupyter environment (PyTorch preconfigured) to run the fine-tuning job directly against internal data, with no server procurement or cluster management.
- Keep training data and checkpoints where they belong. Network Volume provides persistent, shared storage that syncs automatically between Notebook and Inference, so KYC documents, claims data, or clinical text never have to leave a controlled internal path just to move between training and deployment.
- Version and govern what gets deployed. Model Registry tracks every fine-tuned version — imported via Triton, vLLM, or a custom container — with its metadata and lineage, so teams can roll back, compare, or audit which checkpoint is actually serving production traffic.
- Deploy as a low-latency API. Inference turns the fine-tuned model into a RESTful endpoint with autoscaling replica configuration, closing the loop from “fine-tuned model” to “AI feature in the product.”
There’s also a regulatory tailwind pushing enterprises toward the self-hosted lane. Vietnam’s Law No. 134/2025/QH15, effective March 1, 2026, is Southeast Asia’s first binding AI law and includes data residency obligations that push companies training or operating AI systems on Vietnamese data toward local infrastructure — rather than routing sensitive data through third-party APIs whose infrastructure and data handling sit outside a company’s own environment. That’s the same third-party-vs-self-hosted trade-off above, just with a compliance deadline attached. Fine-tuning on a domestic, self-hosted platform isn’t only an accuracy and cost decision anymore — for regulated industries, it’s increasingly a compliance one too. (More on that in our companion piece on why SEA enterprises are bringing LLM customization in-house.)
The real competitive advantage
Using a generic frontier model for every AI task in the business is a bit like using one oversized hammer for every screw in the house — it works often enough to feel productive, right up until the job that actually needed precision. The enterprises closing the accuracy gap in Southeast Asia aren’t necessarily the ones with access to the biggest models. They’re the ones willing to fine-tune a smaller, regionally-grounded model on their own data and own workflows.
That’s a deliberate infrastructure choice, not a lucky prompt. Explore the Model Catalog on GreenNode AI Platform to see which self-hosted base model — and which open regional checkpoint you can bring in via Notebook and Model Registry — gives your team the strongest starting point for fine-tuning on your own domain data.
