Chinese LLM comparison: DeepSeek, Qwen, Kimi, GLM, MiniMax — vendor list prices with dates, context window, thinking mode. In your browser.
Chinese LLM Comparison
Compare the major Chinese chat LLMs side by side — DeepSeek, Qwen (通义千问), Doubao, Kimi, ERNIE (文心), Zhipu GLM and MiniMax — on input price, output price, context window and thinking mode (as the vendor states it). Click any column header to sort. Everything renders in your browser; no model is ever called.
Heads-up: prices were checked against each provider's own price list on the date shown, but providers change them often — the official page prevails (价格已于所示日期对照厂商官方价目表核对，以厂商官方为准). This tool only compares; it never calls a model.
| Model | Provider | Input $/1M | Output $/1M | Context | Thinking |
|---|
Sorting the table down to two or three candidates
Scan the comparison table
The table lists the major Chinese chat models — DeepSeek, Qwen (通义千问), Doubao, Kimi, ERNIE, GLM and MiniMax — with each model's provider, input price, output price, context window and thinking mode (as the vendor states it). The data comes from a shared data module loaded same-origin; everything renders locally in your browser.
Read the price columns
Input and output prices are both in USD per one million tokens. For almost every Chinese model the output price is higher than the input price, so the cost of a long answer is driven mainly by output tokens. Models whose price is not yet confirmed show "— / 待核实" until verified against the provider's own page.
Sort to suit your question
Click any column header to sort: sort input price low-to-high to find the cheapest model, click again to reverse it; or sort by context window to pick a model that fits a long document. Rows with a missing price always sink to the bottom, so they never get in the way.
Shortlist here, confirm on the official page
Use the table to shortlist two or three candidates, then check each provider's official pricing page for the current price and context limit — these prices were checked against each provider's own price list on 25 October 2026, but providers change them often; the provider's official numbers prevail. This tool only compares; it never calls a model.
How the Chinese LLM comparison works
One table for a fast-moving, fragmented market
China's large-language-model market moves fast and is unusually fragmented: DeepSeek, Qwen (通义千问) from Alibaba, Doubao from ByteDance, Kimi from Moonshot, ERNIE (文心) from Baidu, Zhipu's GLM line and MiniMax each ship frequent updates at different price points. Comparing them one provider page at a time is slow and error-prone, because every vendor formats its pricing differently and quotes tokens in its own way. This tool collapses all of that into a single, sortable table: each row shows the model, its provider, the input price, the output price, the context window, and whether it is a reasoning or a general model. The numbers come from a shared data module that the whole RECATOOLS AI-tools family reads, loaded same-origin into your browser, so the comparison renders instantly and stays consistent with our other cost tools.
The most useful first move is to sort. Click the input-price header to rank models cheapest-first and immediately see where DeepSeek and the lighter Qwen tiers sit against the rest; click again to reverse it. Sort by context window instead when your real constraint is fitting a long document or a long conversation into a single request. Models whose price we have not yet confirmed against the provider's own page show "— / 待核实" and always sort to the bottom, so a pending figure never distorts the ranking. Because the table is just a view over structured data, it is honest about what it does and does not know — and it never pretends a missing price is zero.
"Pick a Chinese model the way you'd pick anything else with a price tag: line them up, compare the numbers, then confirm on the official page before you commit."
Input price, output price and context — read all three
Reading the price columns correctly matters. Input price applies to the prompt tokens you send; output price applies to the answer tokens the model writes back. For almost every Chinese model the output price is the larger of the two — often by several times — so a long, generated answer costs far more than a long prompt. Thinking amplifies this: a model that reasons before answering emits a hidden trace first, which means more output tokens and a higher bill, in exchange for stronger performance on hard reasoning, maths and code. GLM-5.3 always does it; DeepSeek V4, Qwen3.8 Max and Qwen3.7 Plus let you switch it off for routine chat, summarisation and drafting. The context-window column is the third number to weigh: it sets how much text — input and output together — a single call can hold, and the verified models span about 262K to 1M tokens.
Treat every figure here as a dated check, not a quote. Each price was read from the provider's own international price list on 25 October 2026; providers change prices, regional list prices and promotional tiers often, and the table can lag them. The right workflow is to use this table to shortlist two or three candidates on price, context and thinking mode, and then open each provider's official pricing page to confirm the current number before you build against it. The tool deliberately does nothing else: it reads our own same-origin data, renders the table locally, and never calls a model or uploads anything you do. That keeps it fast, private and dependable — a starting point for a decision, with the official page always the final word.
One Chinese character is not one token
The Chinese LLM ecosystem is highly diverse: DeepSeek, Qwen (Alibaba), Doubao (ByteDance), Kimi (Moonshot), ERNIE (Baidu), Zhipu GLM and MiniMax each have a distinct niche — this table lines them up for side-by-side comparison.
Input price and output price are two separate numbers: the input price applies to the prompt tokens you send, the output price to the answer tokens the model generates, each billed at its own rate.
For almost every Chinese model the output price is clearly higher than the input price, so the total cost of a long answer is usually driven mainly by output tokens.
When the table was checked on 25 October 2026, the lowest input prices were Qwen3.8 Flash at $0.15 per million tokens, then DeepSeek V4.1 Flash and MiniMax M3 at $0.30.
Thinking costs output tokens: GLM-5.3 always reasons before it answers, while DeepSeek V4, Qwen3.8 Max and Qwen3.7 Plus let you switch thinking on or off — the same question costs more with it on.
The context window sets how many tokens one request can hold (input and output share it): the verified models range from about 262K (Kimi K2.6) to 1M tokens, most current flagships take 1M, and a larger window suits long documents and long conversations.
Token counting for Chinese models differs from English models — one Chinese character is not necessarily one token; the prices shown are per million tokens by each provider's own definition.
Some models (e.g. Doubao, ERNIE) still show "— / 待核实", meaning the price has not yet been confirmed against the provider's official page and will be filled in once verified.
Within one tier the price gap can be several-fold: a side-by-side comparison can cut your API cost markedly while keeping the capability you need.
This tool only reads our own same-origin shared data and renders the table locally in your browser — it never calls DeepSeek, Qwen or any model, and uploads none of your data.
Frequently Asked Questions
- No. It only reads our own same-origin shared model data and renders a comparison table locally in your browser. It does not call DeepSeek, Qwen or any model, and uploads none of your data to any server. It compares information only — no reasoning, no generation.
- It covers DeepSeek (V4 Pro, V4.1 Flash), Qwen (通义千问, Alibaba: 3.8 Max, 3.7 Plus, 3.8 Flash), Kimi (Moonshot: K3, K2.6), Zhipu GLM-5.3 and MiniMax M3, plus Doubao (ByteDance) and ERNIE (Baidu), whose prices are not yet confirmed. As the shared data is updated, added or changed models are reflected automatically.
- All prices are in USD per one million (1M) tokens, with input and output listed separately. For most models the output price is higher than the input price, so the cost of a long answer is driven mainly by output tokens.
- That means the price for that model (e.g. Doubao, ERNIE) has not yet been confirmed against the provider's official pricing page, so no figure is shown. Such rows sort to the bottom automatically and never interfere with comparing the rest.
- Each price was checked against the provider's own international price list on 25 October 2026 — the first (shortest-prompt) tier, and for DeepSeek its peak rate, charged 09:00–12:00 and 14:00–18:00 Singapore time on weekdays except Chinese public holidays; every other hour is half price. Providers change prices often. Always treat the provider's official pricing page as authoritative — use this table to shortlist, then confirm on the official site.
- The context window is the maximum number of tokens a single request can hold, shared between input and output. A larger window fits longer documents, longer conversations or more reference material. Values are shown rounded (e.g. 1M) for easy comparison.
- "Always reasons" means the model writes a thinking trace before every answer (GLM-5.3). "Switchable" means you can turn thinking on or off per request (DeepSeek V4, Qwen3.8 Max and Qwen3.7 Plus). A dash means the vendor's price page does not say. Thinking produces more output tokens, so the same question costs more with it on.
- Yes. Click any column header to sort by that column, and click again to toggle ascending / descending. Common uses are sorting input price to find the cheapest model, or sorting context window to pick one that fits a long document; rows with a missing price always sink to the bottom.
- No. The tool needs no input from you, and the whole table renders locally in your browser. It collects, uploads and stores nothing, and calls no external endpoint (it only loads our own same-origin data file).
- Completely free, with no account or sign-up and no usage limit. It runs in your browser and collects no data.
Method & sources
How it computes
A sortable view of the shared CN chat rate card (AiRateCardKernel): per model, USD per million input and output tokens at the first price tier (DeepSeek at its weekday peak rate), context window, and the thinking mode the vendor states (always / switchable / not stated). Each priced row records the vendor page read and the date. Unconfirmed prices show "— / 待核实" and sort last in both directions.
What this tool implements
- Prices are read from each vendor's own international price list, never an aggregator, and each row records the page and the date. Doubao and ERNIE are not on the rate card yet and are shown as unconfirmed, never converted or guessed.
- Where a vendor tiers by prompt length (Alibaba Qwen3.7 Plus at 256K, MiniMax M3 at 512K) the first tier is shown. DeepSeek's peak rate is shown: its page sets peak at 01:00–04:00 and 06:00–10:00 UTC, Monday to Friday except Chinese public holidays — 09:00–12:00 and 14:00–18:00 in Singapore — and off-peak at half.
- The thinking column states only what the vendor's page says: GLM-5.3 always reasons (Z.ai: disabling reasoning is no longer supported); DeepSeek V4, Qwen3.8 Max and Qwen3.7 Plus are switchable; the rest are left blank.
- Corrected 2026-10-26: Qwen3.7 Plus's thinking mode (shown as not stated; Alibaba prices both modes); the page's check date (14 October, against a table dated 25 October); DeepSeek's peak hours (said to be Singapore working hours, which includes the 12:00–14:00 off-peak gap); a 999,999-token window would have read "1000K".
Sources
- DeepSeek API Docs. Models & Pricing (deepseek-flash = DeepSeek-V4.1-Flash, deepseek-v4-pro; peak hours and off-peak at half), read 2026-10-26: https://api-docs.deepseek.com/quick_start/pricing
- Alibaba Cloud Model Studio. Model inference pricing, International (qwen3.8-max, qwen3.7-plus Non-Thinking and Thinking mode, qwen3.8-flash), last updated 2026-10-22, read 2026-10-26: https://www.alibabacloud.com/help/en/model-studio/model-pricing
- Moonshot AI, Kimi API Platform. Model Pricing (kimi-k3 1,048,576-token window; kimi-k2.6 262,144, K2 series retiring by end of December 2026), read 2026-10-26: https://platform.kimi.ai/docs/pricing/chat
- Z.ai. Pricing (GLM-5.3 $1.4 / $4.4), read 2026-10-26: https://docs.z.ai/guides/overview/pricing
- Z.ai. GLM-5.3 model guide (1M-token context window; always operates with reasoning enabled), read 2026-10-26: https://docs.z.ai/guides/llm/glm-5.3
- MiniMax API Platform. Pricing Overview (MiniMax-M3 ≤512k and >512k input tokens, standard rate after the permanent 50% discount), read 2026-10-26: https://platform.minimax.io/docs/pricing/overview
- MiniMax API Platform. Models (MiniMax-M3: 1M context window), read 2026-10-26: https://platform.minimax.io/docs/guides/models-intro
What can make this go out of date
- The shared rate card public/assets/vendor/reca-ai-data/llm-models.js (chatCN, chatCNAsOf), loaded same-origin. Vendors reprice and release often — on 2026-10-26 Moonshot (kimi-k3.1-pro), Zhipu (GLM-5.3-Flash) and MiniMax (MiniMax-M3.1-Flash) listed models the card does not carry yet — so the rows need re-checking against the pages cited, and the page states the check date.
Pick up where you left off
Stored only in this browser — never sent to our servers.