From OpenAI, Anthropic, Google and xAI to DeepSeek, Qwen, Kimi and GLM — current flagship models, release timelines, context windows, API pricing and whether each can be used inside Edor.ai. Every figure links to official documentation so schools can verify it.
Prices are the vendors' published API rates (USD per 1M tokens) and do not represent what a school spends on Edor.ai — our plans are billed as a per-school annual fee with no token charges.
Prices and specifications in this article are current as of 2026-09
Native provider. OpenAI is the platform default and also supplies the embedding, speech-to-text and text-to-speech models behind the knowledge base and oral practice. Administrators can switch models in the admin panel.
Read the full pageNative provider. Claude is reliable for Chinese writing feedback and long-document work, but it has no embedding or speech models, so the knowledge base and oral practice still need an OpenAI or Azure key.
Read the full pageNot a native provider. Schools can reach the Gemini family through the Poe provider, and Google's open-weight Gemma family can be self-hosted with Ollama. Schools already on Google Workspace for Education can run both side by side.
Read the full pageNot a native provider; reachable through Poe. Grok is tightly integrated with X platform content, so schools should assess its content-safety posture before using it with students.
Read the full pageNot a native provider, but Meta's open-weight models are among the most practical choices for on-premises use: Muse Glimmer 30B ships under Apache 2.0 and runs on a single consumer GPU via Ollama or LM Studio, so student data never leaves the campus network. Note that in April 2026 Meta moved its frontier line from Llama to the closed-weight Muse Spark, making the Llama family legacy.
Read the full pageNot a native provider. Mistral releases its whole line under Apache 2.0 and is the most-cited choice under European data-governance frameworks; the small Ministral 3B / 8B models suit older school hardware or offline use.
Read the full pageNot applicable to a school platform. Composer is a coding-specific model built into the Cursor editor, and the documentation states it is available only inside Cursor's own products, with no inference API that can be called as a model endpoint. Cursor does offer a Cloud Agents API and SDK, but those drive coding agents inside Cursor's environment rather than exposing a chat model a learning platform can call, and it is not designed for classrooms or students. Teachers can use Cursor as a tool when teaching programming, but the platform's model choice belongs among the other providers listed here.
Read the full pageNot a native provider. The practical route is to run a distilled or quantised build on a school server through Ollama so no data leaves the campus network. Calling DeepSeek's own API is a cross-border transfer and needs a privacy assessment first.
Read the full pageNot a native provider. Qwen is the open family most respected for Chinese-language work, and Edor.ai's Ollama suggestions already include `qwen2.5:14b`. The newer Qwen3.8-27B ships under Apache 2.0 and is documented to run in 24GB of VRAM, making it a practical on-premises choice for Chinese material. The flagship Qwen3.8-Max is a closed cloud API and a cross-border transfer, so student data needs assessment first.
Read the full pageNot wired up today. Kimi K3 publishes weights, but at 2.8T parameters (about 104B active) it is far beyond what a school server can host, so self-hosting is unrealistic. Its official API sits in mainland China and is a cross-border transfer. Schools wanting to try its long-document strength should have teachers use a personal account on de-identified material only.
Read the full pageNot a native provider. GLM is one of the few teams publishing large models under MIT (GLM-5.2 and GLM-5.3-Flash), which appeals to sponsoring bodies wanting full autonomy, though the parameter counts still need server-class hardware and weight releases do not land at the same time for every version — verify against the official Hugging Face page before committing.
Read the full pageNot wired up today. MiniMax M3 takes a sparse-attention, low-cost long-context approach — about 428B total parameters with only ~23B active per token — making it one of the cheapest long-context models in its class. Its API is in mainland China and its licence terms need line-by-line review, so an IT coordinator should assess privacy and licensing first.
Read the full pageNot wired up, and not recommended for student-facing use in Hong Kong schools. Doubao is served only through Volcano Engine in mainland China, its weights are closed, its pricing is in RMB (Seed 2.1 Pro at ¥6 input / ¥30 output per 1M tokens), and its terms and data location sit under mainland regulation. Where Chinese-language strength is the goal, self-hosting Qwen or GLM open weights is the more controllable route.
Read the full pageNot a native provider, but notable: Hunyuan Hy4 Preview publishes 770B-parameter weights (49B active) under a fully permissive Apache 2.0 licence, which is clearer than the bespoke licences on GLM-5.3 or Kimi K3. Self-hosting still needs server-class hardware, and calling Tencent Cloud's API directly is a cross-border transfer.
Read the full pageNot wired up today. ERNIE's main strength is search-grounded answers from Baidu, which is precisely where Hong Kong schools should be careful: its sources, content policy and data handling all sit in the mainland framework. Its 128K context window is also well below other current flagships.
Read the full pageRecent public releases across vendors, newest first. Generations turn over quickly, so ask any vendor whether model upgrades are already included in the annual fee.