Choosing LLMs in Practice: Balancing Capability, Speed, and Cost
This article outlines a practical way to choose large language models for development work. It explains how to tier models by capability and cost, when multimodal models justify their overhead compared with text-only models, and a five-step decision framework based on input type, task criticality, latency, data compliance, and A/B testing on your own workloads.
A Rational Methodology for LLM Selection
1. What the Release Frenzy Tells Us
Over the past year, the large language model (LLM) industry has entered a period of dense releases. Chinese vendors and leading teams abroad each launch a new generation of models roughly every one to two months. Some emphasize overall capability gains, others focus on reasoning depth, and others highlight multimodal understanding. Among overseas players, the leading models have uniformly moved toward an all-modality approach, integrating text, image, and audio and video understanding into a single product.
Chinese vendors show a different landscape. Many teams organize their models by capability. Some stay with a text-only approach and put their parameters and compute entirely into language understanding and generation quality. Others move toward multimodal designs that support image understanding and even video parsing. With so many options, developers face a practical question: which model should they use day to day?
The author’s conclusion is to use Chinese models for most daily tasks and to rely on a small number of leading overseas models to fill capability gaps. This conclusion comes from a change in hands-on experience, not from brand preference. About a year ago, the author’s workflow relied mainly on overseas models. Over the past year, however, Chinese models have become stable enough for mainstream tasks such as code generation, document writing, and logical analysis. In the author’s experience, they can handle more than 90% of daily work.
2. Capability Tiers: Why “Primary Plus Backup”
Splitting models into “Chinese” and “overseas” is not a scientifically useful distinction. A better approach is to tier models by capability ceiling and usage cost.
Tier 1: Daily primary models. These models handle the most frequent tasks: polishing text, completing code, organizing materials, and troubleshooting problems. For this tier, a model does not need to be the strongest at every task. It needs a stable overall experience: fast responses, a long enough context window, natural Chinese and English expression, and few basic mistakes. Leading Chinese text-only models already meet these requirements.
Tier 2: Backup for critical tasks. Architecture design reviews, analysis of complex technical trade-offs, and final review of long documents carry a high cost when they go wrong. These tasks justify a model with a higher capability ceiling. At this tier, the author mainly uses two overseas models. One excels in overall capability and suits architecture suggestions and reviews for critical code. The other performs strongly in multimodal understanding and can directly recognize video content. In scenarios that involve only image or video material, it is an irreplaceable choice.
Tier 3: Vertical specialization. For tasks such as tabular data analysis, mathematical reasoning, and domain-specific knowledge Q&A, choose a model on demand based on its strength in that dimension, rather than looking for an all-rounder.
The core idea of this tiered approach is to let the most expensive models handle only the most expensive tasks. Model calls carry both latency and monetary costs. Routing every request to the most capable model wastes resources.
3. Multimodal vs. Text-Only: Two Technical Paths
A clear split among Chinese model vendors is whether their models support multimodal input.
Multimodal models can read screenshots, design drafts, and scanned documents directly. Some can also understand sequences of video frames. This matters for tasks that involve visual material, such as debugging front-end rendering issues, interpreting architecture diagrams, and analyzing error screenshots. Text-only models require the user to describe the image content in words. This is inefficient and can lose important details.
Text-only vendors took a different path. They concentrate all training and inference resources on the language model itself. This brings two direct advantages.
The first is speed. Text-only models do not need a visual encoder, so their inference pipeline is shorter. In long conversations and high-concurrency scenarios, the difference in response speed becomes more noticeable. In the author’s tests, the response latency of some multimodal flagship models was several times that of a leading text-only model. For developers who want smooth interaction, this difference can influence daily choices.
The second is more controllable cost and stability. Without visual capability, a model has a more focused optimization space.
The author has used one Chinese text-only model as the primary model for this reason. Its overall capability is strong enough, and its speed is much better than that of multimodal flagship models. For research, development, and writing work centered on text, the text-only path currently offers better value for money.
Why does this split exist? A plausible explanation is compute constraints. Multimodal models need an additional visual encoding module, which significantly increases compute for both training and inference. With high-end compute supply tight, vendors must choose between full multimodal capability and peak performance and speed in a single dimension. This also explains why the release cycles of the text-only and multimodal camps often compete directly. Each path needs to show its pace of progress at the same time, and both aim at the same developers, who are sensitive to price and performance.
4. A Practical Model Selection Framework
Based on the analysis above, you can follow this simplified decision process:
- Determine the input modality first. Does the task involve images, screenshots, or video? If so, choose a multimodal model. If the task involves only text, move to the next step.
- Assess how critical the task is. Is it a key architecture decision, a production code review, or everyday Q&A and drafting? Send the first to the model with the highest capability ceiling, and send the last to your daily primary model.
- Assess latency sensitivity. Interactive coding assistance and real-time conversation should favor faster text-only models. Offline batch tasks can accept slower but more thorough models.
- Evaluate data and compliance requirements. For scenarios involving sensitive enterprise data, prioritize services that support private deployment or offer enterprise-grade data isolation. This constraint often matters more than model capability rankings.
- Build a habit of A/B testing. Run two or three candidate models on the same set of representative tasks, and let your own business data decide. Do not rely only on public leaderboards. Leaderboards reflect general capability, while your scenario may have its own distribution of tasks.
5. Closing Thoughts
No single standard answer exists for LLM selection, but there is a standard set of questions. What is the input modality? How costly is an error? How much latency can you tolerate? Can the data leave your organization’s boundary? Answering these four questions usually narrows the candidates to one or two models.
Chinese models already have enough capability for most development work on text tasks. Using leading overseas models as a backup for multimodal and critical tasks is a practical and economical approach. The two technical paths will likely converge as the industry evolves. While compute constraints persist, however, understanding the technical logic behind each vendor’s trade-offs will help you make more stable choices than chasing every new release.
