How easy-learn-ai Reorganized an AI Model Database into 19 Vendor Collections

Forum topic · 小凯 · 2026-08-17

Summary

Commit e6c189a in the easy-learn-ai project reorganized a single database containing more than 5,000 lines of AI model data into 19 vendor-specific JSON files. The change separates records for organizations such as OpenAI, Google, Alibaba, DeepSeek, and Anthropic, making model updates easier to review, compare, and maintain in parallel. Each collection retains structured information including release dates, capability labels, context and output limits, pricing details, and links to related papers. The resulting database documents nearly 100 models and illustrates several industry directions: million-token context windows, thinking and non-thinking modes, mixture-of-experts architectures, multimodal input support, and model distillation. The redesign also allows image, video, and language model families to evolve independently while following a common data format. More than a code refactor, the commit turns a difficult-to-search reference into a scalable map of the rapidly expanding AI ecosystem, helping developers and learners identify models, compare capabilities, and choose suitable options for different applications.

From a 5,000-page book to 19 shelves

Commit e6c189a in easy-learn-ai reorganized a single, dense model database into 19 vendor-specific JSON files. Each file acts as an independent collection for a provider such as OpenAI, Google, Alibaba, DeepSeek, or Anthropic.

The previous structure can be imagined as one 5,000-page catalog containing every model and its attributes. Finding a specific entry required navigating a large shared file, while contributors updating different providers could encounter overlapping changes and merge conflicts. The new layout functions more like a museum with clearly labeled shelves.

Model records include information such as release dates, capability tags, context-window and output limits, and links to related papers where available.

Why a multi-file architecture matters

This was not merely a file-splitting exercise. It changed the database from a centralized structure into a federated one:

Parallel maintenance: A Qwen update does not require editing the OpenAI collection.
Independent evolution: Different model types can support specialized fields, such as resolution data for image models or context length for language models.
Clear ownership: Problems can be traced to a specific provider’s data file.
Lower contribution barriers: Contributors can work within a smaller, more understandable scope.
Scalable organization: New vendors and models can be added without further enlarging one monolithic document.

The refactor therefore prepares the project for continued growth in the number of models, providers, modalities, and evaluation criteria.

The database as an AI industry map

Reading the 19 collections reveals several distinct strategies among major model providers.

United States providers

OpenAI’s collection is the largest, covering a progression from GPT-3.5 Turbo through the GPT-4.1 family. The range reflects a tiered product strategy: flagship models target maximum capability, general-purpose models balance performance and cost, and smaller versions address lower-cost use cases.

Google’s Gemini series emphasizes long context, with versions such as Gemini 2.5 Pro and 3.5 Flash listed as supporting 1M-token context windows. Large context can be useful for analyzing books, legal documents, and entire codebases in a single request.

Anthropic’s Claude collection emphasizes deep reasoning. Claude Opus 4.8 is listed with 1M-token context, 128K output, and adaptive thinking, allowing the model to adjust its reasoning effort according to task difficulty.

xAI’s Grok entries document a rapid release progression, from the open-weight Grok-1, listed with 314B parameters, to Grok 4. The data shows release timing, although it cannot independently establish how Grok compares with competing models on quality.

Chinese providers

Alibaba’s Qwen family demonstrates a combination of rapid iteration, open-weight releases, proprietary services, and competitive pricing. The post highlights Qwen3.5-Plus as a mixture-of-experts model with 397B total parameters and 17B activated parameters. It also describes later Qwen generations with 1M-token context, large output limits, and thinking budgets.

DeepSeek’s R1 series represents a different strategy centered on open access and distillation. The post lists seven distilled versions ranging from 1.5B to 70B parameters, allowing organizations with suitable hardware to run smaller reasoning models locally instead of relying only on costly APIs.

Other Chinese providers documented in the database include Moonshot AI and Kimi, Zhipu AI and GLM, Baidu and ERNIE, Tencent and Hunyuan, and ByteDance and Seed. Kimi is presented as an example of a long-context and multimodal product strategy.

European and specialist providers

Germany’s Black Forest Labs is represented by the FLUX image-generation family. The post describes FLUX.2 [pro] as a Rectified Flow Transformer-based model supporting multiple reference images, JSON-based structured control, and output resolutions up to 4MP.

The database also includes specialist creative-AI providers such as Stability AI, Midjourney, Runway, and Pika. Their presence shows that image and video generation has developed into a substantial segment of the AI industry rather than remaining a single-model category.

Five technical trends visible in the data

1. The context-window race

The post describes a shift from 8K-token context being common in 2023 to 1M-token context appearing across flagship offerings from several providers. Long context can support repository-level code analysis, full legal-document review, and processing an entire book at once.

2. Thinking modes are becoming standard

Models associated with explicit reasoning include DeepSeek-R1, Qwen3-Thinking, and Gemini 3.1 Pro. Separate fast and slow thinking modes let systems favor quick responses for routine interactions while allocating more computation to mathematics, coding, and logical reasoning.

3. Mixture of Experts reduces activation cost

A mixture-of-experts model contains multiple specialized components but activates only a subset for each input. The post uses Qwen3.5-Plus as an example, with 397B total parameters but 17B activated parameters. This approach aims to retain broad model capacity while reducing the computation required per token.

4. Multimodality is becoming an entry requirement

Newer flagship models increasingly accept combinations of text, images, audio, video, and documents. Google’s Gemini family is cited as supporting multiple input forms. As a result, the database treats multimodal capability as a central comparison point rather than an optional feature.

5. Distillation makes advanced capabilities more accessible

Distilled R1 models allow smaller parameter counts to inherit reasoning behavior from a larger model. The post highlights versions as small as 1.5B parameters, which may be suitable for local deployment on appropriately equipped devices. This challenges the assumption that quality must always scale directly with parameter count.

What easy-learn-ai contributes

easy-learn-ai does not train or host the models listed in its database. Instead, it provides a navigable model map. Developers and learners can use the collections to answer practical questions such as:

Which providers and models are available?
What context, output, and modality limits apply?
Is a model open-weight or proprietary?
What capabilities and pricing information are documented?
Which model best fits a particular application?

Moving from one monolithic file to 19 structured collections is therefore both a maintenance improvement and a response to the AI industry’s rapid expansion. As providers multiply and model specifications change, consistent organization becomes essential for keeping the knowledge useful, searchable, and maintainable.