DeepSeek, Qwen, Kimi and GLM are the four Chinese model families most often compared with each other, yet their providers publish prices, limits, licences and benchmarks in different formats, so a fair side by side view takes work. This page sets the current flagship and the current fast tier of each family next to each other using identical fields: API identifier, parameters, context window, output limit, input types, reasoning controls, list price, open weights, licence, availability, provider-reported benchmarks and deprecations. The providers are DeepSeek (DeepSeek), Z.ai, formerly Zhipu (GLM), Moonshot AI (Kimi) and Alibaba Cloud (Qwen). Families are listed alphabetically by family name in every table: DeepSeek, GLM, Kimi, Qwen; order is never a ranking. It is written for developers, architects, procurement and governance teams who need checked numbers. Where a provider does not publish a fact on the pages checked, the table says so rather than estimating.
Last checked: 2026-11-15.
This is an independent reference page from NHI Mgmt Group based on public provider documentation; it is not sponsored or endorsed by any provider. Model details change quickly, so confirm figures against each provider's current documentation, linked under Standards and references.
Key takeaways
- Eight models, four families. DeepSeek-V4-Pro and DeepSeek-V4.1-Flash, GLM-5.3 and GLM-5.3-Flash, Kimi K3.1 Pro and Kimi K3, Qwen3.8-Max and Qwen3.8-Flash, each as served on the provider's own API on 15 November 2026.
- Every flagship has downloadable weights. DeepSeek publishes its weights under the MIT License; Moonshot AI, Alibaba and Z.ai use their own model licences for their flagships, while GLM-5.3-Flash is MIT licensed.
- Around 1M tokens of context at the top. DeepSeek, Z.ai and Moonshot AI state 1M or 1,048,576 tokens for their flagships; Alibaba states 262,144 tokens natively for the open Qwen3.8 weights, extensible to 1,010,000, with a 1M default on the hosted Qwen3.8-Max.
- Pricing levers differ. DeepSeek halves prices outside its peak windows, Alibaba prices by deployment region, Moonshot AI bills batch at 60% of standard, and Z.ai lists cached-input storage as limited-time free.
- Reasoning is on by default in all four. Each flagship exposes an effort setting (low, high and max, or low, medium and xhigh for Qwen), and GLM-5.3 and the open Qwen3.8 weights cannot switch reasoning off.
Which models are compared?
For each family this page takes the model the provider's documentation positions at the top of its current line, plus the current lower-cost tier, as of 15 November 2026. Older generations are covered in each family's version comparison and provider hub pages, linked below, and the wider cross-lab view is in Frontier AI Models Compared. The China and US flagship view is in Chinese vs US Frontier Models Compared; this page goes one level deeper on these four families only.
| Family | Provider | Flagship covered | Fast tier covered | Basis for selection | Source |
|---|---|---|---|---|---|
| DeepSeek | DeepSeek | DeepSeek-V4-Pro-0813 (deepseek-v4-pro) | DeepSeek-V4.1-Flash-1017 (deepseek-flash) | The two models on DeepSeek's API pricing page | DeepSeek pricing |
| GLM | Z.ai | GLM-5.3 (glm-5.3) | GLM-5.3-Flash | Z.ai calls GLM-5.3 "Z.ai's latest flagship model" | Z.ai GLM-5.3 |
| Kimi | Moonshot AI | Kimi K3.1 Pro (kimi-k3.1-pro) | Kimi K3 (kimi-k3) | The two K3-generation models on the Kimi platform pricing page | Kimi pricing |
| Qwen | Alibaba Cloud | Qwen3.8-Max (qwen3.8-max) | Qwen3.8-Flash (qwen3.8-flash) | Max is "the official version based on Qwen3.8-2.4T-A95B"; Flash is listed as "Fast and cost-effective" | Qwen3.8 model card |
Kimi K3 shares the parameter count and context window of K3.1 Pro; it fills the second Kimi row because Moonshot AI lists no smaller K3-generation model on the pages checked.
DeepSeek, Qwen, Kimi and GLM at a glance
As of 2026-11-15. "Not stated" means the provider did not publish the fact on the pages checked (2026-11-15). Parameter counts are as shown on each Hugging Face model card.
| Model | API ID | Total / active parameters | Context window | Max output | Input types | Source |
|---|---|---|---|---|---|---|
| DeepSeek-V4-Pro-0813 | deepseek-v4-pro | 1.7T / not stated by DeepSeek | 1M | 384K | Text | Model card, pricing |
| DeepSeek-V4.1-Flash-1017 | deepseek-flash | 763B / not stated by DeepSeek | 1M | 384K | Text, image | Model card, pricing |
| GLM-5.3 | glm-5.3 | 753B / not stated by Z.ai | 1M | 128K | Text only | Z.ai docs, model card |
| GLM-5.3-Flash | GLM-5.3-Flash | 320B / 18B | Not stated by Z.ai | Not stated by Z.ai | Text, image | Model card |
| Kimi K3.1 Pro | kimi-k3.1-pro | 2.8T / 104B | 1,048,576 | Not stated by Moonshot AI | Text, image | Model card |
| Kimi K3 | kimi-k3 | 2.8T / 104B | 1,048,576 | Not stated by Moonshot AI | Text, image (card text also mentions video) | Model card |
| Qwen3.8-Max | qwen3.8-max | 2.4T / 95B (open weights Qwen3.8-2.4T-A95B) | 1M default on the hosted model; open weights 262,144 native, up to 1,010,000 | Not stated by Alibaba | Text, image (hosted); text only (open weights) | Model card |
| Qwen3.8-Flash | qwen3.8-flash | Not stated by Alibaba | Not stated by Alibaba | Not stated by Alibaba | Text | Model Studio models |
The Kimi K3 card lists text and image input in its specification table but its introduction also mentions video, so confirm video support with Moonshot AI before relying on it. Model-by-model detail is in DeepSeek V4.1 Flash, Kimi K3 and K3.1 and Qwen3.8 (Max and 27B).
When was each model released?
Release facts as each provider states them, as of 2026-11-15. Full histories are in the LLM Release Tracker.
| Model | Release facts stated by the provider | Source |
|---|---|---|
| DeepSeek-V4-Pro-0813 | V4-Pro and V4-Flash added on 24 April 2026; general availability version of V4-Pro released on 13 August 2026 | DeepSeek change log |
| DeepSeek-V4.1-Flash-1017 | V4.1-Flash released on 10 September 2026; upgraded to the 1017 version on 17 October 2026 with unchanged calling method and pricing | DeepSeek change log |
| GLM-5.3 | Release date not stated by Z.ai on the pages checked; the docs say it is available to all GLM Coding Plan users, and it uses the same base model as GLM-5.2 | Z.ai GLM-5.3 |
| GLM-5.3-Flash | Release date not stated by Z.ai on the pages checked; described as the first natively multimodal model in the GLM-5 series | Model card |
| Kimi K3.1 Pro | Release date not stated by Moonshot AI on the pages checked | Model card |
| Kimi K3 | Release date not stated by Moonshot AI on the pages checked | Model card |
| Qwen3.8-Max | Model card citation dated August 2026; no separate release date stated | Model card |
| Qwen3.8-Flash | Release date not stated by Alibaba on the pages checked | Model Studio models |
How much do DeepSeek, Qwen, Kimi and GLM cost?
Standard list prices per million tokens on each provider's own API, as of 2026-11-15. DeepSeek and Alibaba show prices with a "$" sign; Z.ai states USD. Alibaba prices shown are the Singapore (International) deployment.
| Model | Input | Cached input | Output | Notes | Source |
|---|---|---|---|---|---|
| DeepSeek-V4-Pro | $1.32 peak / $0.66 off-peak | $0.044 peak / $0.022 off-peak | $3.96 peak / $1.98 off-peak | Concurrency limit 500 | DeepSeek pricing |
| DeepSeek-V4.1-Flash | $0.30 peak / $0.15 off-peak | $0.006 peak / $0.003 off-peak | $1.20 peak / $0.60 off-peak | Concurrency limit 2,500 | DeepSeek pricing |
| GLM-5.3 | $1.40 | $0.26 | $4.40 | Cached input storage limited-time free | Z.ai pricing |
| GLM-5.3-Flash | $0.15 | $0.03 | $0.50 | Cached input storage limited-time free | Z.ai pricing |
| Kimi K3.1 Pro | Not stated by Moonshot AI on the pages checked | Not stated by Moonshot AI on the pages checked | Not stated by Moonshot AI on the pages checked | Batch billed at 60% of the standard price | Kimi pricing |
| Kimi K3 | $3.00 | $0.30 | $15.00 | Launch prices from the K3 blog; batch billed at 60% of the standard price | Kimi K3 blog, Kimi pricing |
| Qwen3.8-Max | $2 | Not stated per model by Alibaba | $6 | Single tier up to 1M tokens; free quota of 1M tokens | Model Studio pricing |
| Qwen3.8-Flash | $0.15 | Not stated per model by Alibaba | $0.47 | Single tier up to 1M tokens; free quota of 1M tokens | Model Studio pricing |
Kimi K3 prices come from Moonshot AI's K3 launch blog. The Kimi platform pricing tables are rendered by script and their figures did not appear in the page text retrieved on 15 November 2026; read them directly on the Kimi platform before budgeting. Current rates for every provider are also collected in the LLM Pricing Tracker.
Pricing structures
- DeepSeek peak windows: peak hours are 01:00 to 04:00 and 06:00 to 10:00 UTC, Monday to Friday, excluding Chinese public holidays; all other hours are off-peak at half the peak rate. Peak and off-peak pricing took effect at 16:00 UTC on 16 August 2026, per the change log.
- Z.ai storage and free models: cached input storage is listed as limited-time free for the paid GLM-5 models, and GLM-4.7-Flash, GLM-4.5-Flash and GLM-4.6V-Flash are listed as free in every column. No batch price is listed.
- Moonshot AI batch: kimi-k3.1-pro and kimi-k3 support batch inference billed at 60% of the standard price.
- Alibaba deployment scope: qwen3.8-max costs $2 input and $6 output in Singapore (International) and US (Virginia) deployment, and $1.65 input and $4.951 output in China (Beijing) and the Global scope (Hong Kong, Germany, Japan, US Virginia). A 50% batch discount is listed for the China (Beijing) row only.
What is the context window of each model?
| Model | Context window | Max output | Notes | Source |
|---|---|---|---|---|
| DeepSeek-V4-Pro | 1M | 384K | The model card recommends a 384K maximum output length for the high and max reasoning levels | Model card |
| DeepSeek-V4.1-Flash | 1M | 384K | Output limit is listed once on the pricing page for both models | DeepSeek pricing |
| GLM-5.3 | 1M | 128K | Text-only input | Z.ai GLM-5.3 |
| GLM-5.3-Flash | Not stated by Z.ai | Not stated by Z.ai | 300,000 tokens appears only as a benchmark evaluation setting | Model card |
| Kimi K3.1 Pro | 1,048,576 | Not stated by Moonshot AI | Described as a 1-million-token context window | Model card |
| Kimi K3 | 1,048,576 | Not stated by Moonshot AI | Same window as K3.1 Pro | Model card |
| Qwen3.8-Max | 1M default (hosted) | Not stated by Alibaba | Open weights: 262,144 native, extensible to 1,010,000 | Model card |
| Qwen3.8-Flash | Not stated by Alibaba | Not stated by Alibaba | Pricing tier runs to 1M input tokens | Model Studio pricing |
Reasoning modes and API features
All four flagships are reasoning models with an effort parameter. The settings and API features below are as each provider documents them, as of 2026-11-15.
| Family | Reasoning control | Can reasoning be turned off? | Documented API features | Source |
|---|---|---|---|---|
| DeepSeek | Three thinking effort levels: low, high and max (added 13 August 2026) | Yes: non-thinking and thinking modes, thinking is the default | JSON output, tool calls, FIM completion (beta, non-thinking only), Chat Prefix Completion (beta), Responses API, Anthropic API format; vision on deepseek-flash only | DeepSeek pricing, change log |
| GLM | reasoning_effort: low, high, max (default max) | No for GLM-5.3: "Disabling reasoning is no longer supported" | Function calling, structured output, context caching, streaming; GLM-5.3-Flash adds a clear_thinking setting | Z.ai GLM-5.3, GLM-5.3-Flash card |
| Kimi | reasoning_effort: low, high, max (default max) | Not stated by Moonshot AI on the pages checked | OpenAI-compatible and Anthropic-compatible endpoints; batch inference | Model card, Kimi pricing |
| Qwen | reasoning_effort: low, medium, xhigh (default xhigh); preserve_thinking on by default | Open weights: no ("thinking cannot be disabled"); hosted Qwen3.8-Max adds non-thinking support | Hosted Max adds vision input, a 1M default context and official built-in tools not in the open weights | Model card |
DeepSeek also opened its in-house search engine through an API on 17 October 2026, and on 13 November 2026 added web search, file input, URL reading and model safety to deepseek-flash, plus standalone Search and Fetch, OCR, ASR and TTS APIs.
Are DeepSeek, Qwen, Kimi and GLM open source?
Seven of the eight models have weights on Hugging Face under the provider's organisation; for the qwen3.8-flash API model, Alibaba does not state published weights on the pages checked. Open weights are not the same as an open-source licence: read the licence file before commercial use, fine-tuning or redistribution. For background see Open-Weight vs Closed Models Explained.
| Model | Weights published | Licence named on the model card | Inference engines named | Source |
|---|---|---|---|---|
| DeepSeek-V4-Pro-0813 | Yes | MIT License (repository and weights) | vLLM, SGLang, Transformers, Docker Model Runner | Model card |
| DeepSeek-V4.1-Flash-1017 | Yes | MIT License (repository and weights) | vLLM, SGLang, Transformers, Docker Model Runner | Model card |
| GLM-5.3 | Yes | glm-5.3 (custom) | SGLang, vLLM, TokenSpeed, Transformers, KTransformers, Unsloth, Ascend NPU options | Model card |
| GLM-5.3-Flash | Yes | MIT | SGLang, vLLM, TokenSpeed, Transformers, KTransformers, Unsloth | Model card |
| Kimi K3.1 Pro | Yes | Kimi K3.1 Pro License (custom) | vLLM, SGLang, TokenSpeed (recommended), Transformers | Model card |
| Kimi K3 | Yes | Kimi K3 License (custom) | vLLM, SGLang, TokenSpeed (recommended) | Model card |
| Qwen3.8-Max | Yes, as Qwen3.8-2.4T-A95B (without the hosted extras) | qwen3.8-max (custom) | SGLang, vLLM, TokenSpeed | Model card |
| Qwen3.8-Flash | Not stated by Alibaba for the API model | Not stated by Alibaba | Not stated by Alibaba | Model Studio models |
Self-hosting the flagships needs data-centre hardware: DeepSeek's V4-Pro card gives a vLLM example on a single node of four GB300 accelerators, and Kimi K3.1 Pro ships MXFP4 weights produced with quantisation-aware training. Hardware sizing across open models is covered in Open-Weight Frontier Models Compared.
Where are the models available?
| Family | Provider API | Weights and third-party hosting named by the provider | Source |
|---|---|---|---|
| DeepSeek | DeepSeek API (OpenAI Chat Completions, Anthropic and Responses formats, same base URL) | Hugging Face; Hugging Face inference providers including DeepInfra; community quantisations for llama.cpp, Ollama and LM Studio | Change log, model card |
| GLM | Z.ai API; GLM Coding Plan | Hugging Face; Hugging Face inference providers including DeepInfra; quantised builds for llama.cpp, Ollama and LM Studio | Z.ai GLM-5.3, model card |
| Kimi | Kimi platform (platform.kimi.ai), OpenAI-compatible and Anthropic-compatible | Hugging Face; GGUF quantisations for llama.cpp, Ollama and LM Studio | Model card |
| Qwen | Alibaba Cloud Model Studio, with Singapore (International), China (Beijing), Global and US (Virginia) deployment scopes | Hugging Face (open weights); quantisations for llama.cpp, Ollama and LM Studio | Model Studio pricing, model card |
Provider hubs list every model and channel for each family: DeepSeek Models, Zhipu Z.ai GLM Models, Moonshot AI Kimi Models and Alibaba Qwen Models.
Benchmarks
Provider-reported scores only, from each model card or documentation page, as of 2026-11-15. Two benchmarks are shown because they are the ones most of these providers report: DeepSWE v1.1 and Terminal-Bench 2.1. Providers run them with their own harnesses and effort settings, so a difference of a few points is not a like-for-like result. No independent leaderboard figures are included. How to read such tables is covered in How to Read LLM Benchmarks.
| Model | DeepSWE v1.1 | Terminal-Bench 2.1 | Notes | Source |
|---|---|---|---|---|
| DeepSeek-V4-Pro-0813 | 62.7 (card says "DeepSWE", version not stated) | 87.9 | Self-reported on the model card | Model card |
| DeepSeek-V4.1-Flash-1017 | 77.0 | Not stated by DeepSeek (card reports Terminal-Bench 4.0: 42.3) | Harness named as Mini-SWE-V2 for DeepSWE | Model card |
| GLM-5.3 | 66.9 | Not stated by Z.ai (docs report Terminal-Bench 3.0) | Reported as an improvement from GLM-5.2's 46.2 | Z.ai GLM-5.3 |
| GLM-5.3-Flash | Not stated by Z.ai on the pages checked | Not stated by Z.ai on the pages checked | Model card | |
| Kimi K3.1 Pro | 75.5 | Not stated by Moonshot AI (card reports Terminal-Bench 4.0: 65.7) | Max effort | Model card |
| Kimi K3 | 67.5 | 88.3 | The card notes 67.3 on the official DeepSWE leaderboard with the mini-SWE-agent harness; Terminal-Bench run with the Kimi Code harness | Model card |
| Qwen3.8-Max (open weights) | 56.6 | 86.6 | Scores are for Qwen3.8-2.4T-A95B | Model card |
| Qwen3.8-Flash | Not stated by Alibaba on the pages checked | Not stated by Alibaba on the pages checked | Model Studio models |
Deprecations and retirements
| Family | Model or item | Details | Source |
|---|---|---|---|
| DeepSeek | deepseek-chat, deepseek-reasoner | Discontinued on 24 July 2026; until then they mapped to V4-Flash non-thinking and thinking modes | Change log |
| DeepSeek | deepseek-v4-flash, deepseek-v4-flash-vision-exp | Models retired on 10 September 2026; the names are still accepted and served by DeepSeek-V4.1-Flash-1017 at the Flash price | Pricing |
| GLM | GLM-5.2, GLM-5.1, GLM-5 | No retirement date stated by Z.ai; all three still priced on 15 November 2026 | Z.ai pricing |
| Kimi | K2 series | Not available to newly registered users from 31 October 2026; retirement of all K2 series models planned by the end of December 2026 | Kimi pricing |
| Qwen | qwen3.7-max, qwen3.7-plus | No retirement date stated by Alibaba on the pages checked; both still priced on 15 November 2026 | Model Studio pricing |
DeepSeek's routing of retired names means an application can move to a new model without a code change. Record the model version and re-test after each update. Dates across all providers are tracked in the LLM Deprecation and Retirement Tracker.
Data handling and security
- Hosted API or self-hosted: all four families can be called on the provider's API or run from published weights. With self-hosting, prompts stay in your environment but you own patching, access control, logging and provenance checks; see the Self-Hosted LLM Security Guide and the Open-Weight Model Security Guide.
- Processing location: Alibaba documents distinct deployment scopes, including China (Beijing) and US (Virginia). DeepSeek, Moonshot AI and Z.ai do not state processing regions on the pages checked; read each provider's terms and privacy policy for the plan you buy.
- Compatible endpoints: DeepSeek and Kimi accept OpenAI-format and Anthropic-format requests, so one base URL change can send an existing agent's traffic, including code and secrets in context, to a different provider. Control that setting centrally.
- API keys are non-human identities: store them in a secrets manager, scope them per application and set spend limits. See the LLMjacking Guide.
- Policy restrictions: some organisations restrict particular providers or jurisdictions. The assessment approach is in the DeepSeek and High-Risk AI Vendor Assessment Guide, and retention questions are in AI Data Retention and Deletion.
How to choose between DeepSeek, Qwen, Kimi and GLM: factual criteria
These are criteria, not recommendations. Score each candidate against your own workload, policy and evaluation results.
| Criterion | Question to answer | Where to look on this page |
|---|---|---|
| Licence | Does MIT or a custom model licence fit your commercial use, fine-tuning and redistribution? | Open source table; the licence file itself |
| Input types | Do you need image or video input, and on which model? | At a glance table |
| Reasoning control | Do you need a non-thinking mode for low-latency calls? | Reasoning modes table |
| Context and output | Do your documents exceed 262,144 tokens, and do you need outputs above 128K? | Context window table |
| Cost at your volume | What is the blended price with caching, off-peak windows, batch and region? | Pricing tables |
| Processing location | Must data be processed in a particular region? | Availability and data handling sections |
| Change cadence | How often does the provider replace models, and does it route old names? | Deprecations table |
| Self-hosting capacity | Do you have the accelerators to serve a model with hundreds of billions to trillions of parameters? | At a glance and open source tables |
Small teams usually start on a hosted API with published prices. Larger organisations with sovereignty, sector or procurement rules tend to weigh licence terms, processing location and self-hosting more heavily, and should run their own evaluations rather than rely on provider-reported scores.
Practitioner checklist
- Record the exact model version for each integration (for example DeepSeek-V4.1-Flash-1017 behind deepseek-flash), not only the API alias.
- Read the licence file for the specific weights before commercial use; MIT and custom model licences carry different terms.
- Price your workload with caching, DeepSeek off-peak hours, Kimi batch and Alibaba deployment scope applied, and confirm Kimi prices directly on its platform.
- Check whether reasoning can be disabled for the model you choose before designing latency-sensitive paths.
- Run your own evaluation set on every shortlisted model; treat provider-reported benchmarks as indicative only.
- Control base URL and endpoint settings centrally so agent traffic cannot be redirected to another provider without approval.
- Store each provider's API keys in a secrets manager with per-application scope, rotation and spend caps.
- Diarise known retirements, such as the Kimi K2 series by the end of December 2026, and re-test when providers route retired names to new models.
Standards and references
- DeepSeek: Models and pricing (accessed 2026-11-15)
- DeepSeek: API change log (accessed 2026-11-15)
- DeepSeek: DeepSeek-V4-Pro-0813 model card (accessed 2026-11-15)
- DeepSeek: DeepSeek-V4.1-Flash-1017 model card (accessed 2026-11-15)
- Z.ai: GLM-5.3 documentation (accessed 2026-11-15)
- Z.ai: Pricing (accessed 2026-11-15)
- Z.ai: GLM-5.3 model card (accessed 2026-11-15)
- Z.ai: GLM-5.3-Flash model card (accessed 2026-11-15)
- Moonshot AI: Kimi platform chat pricing (accessed 2026-11-15)
- Moonshot AI: Kimi K3.1 Pro model card (accessed 2026-11-15)
- Moonshot AI: Kimi K3 model card (accessed 2026-11-15)
- Moonshot AI: Introducing Kimi K3 (accessed 2026-11-15)
- Alibaba Cloud: Model Studio model pricing (accessed 2026-11-15)
- Alibaba Cloud: Model Studio models (accessed 2026-11-15)
- Alibaba Qwen: Qwen3.8-2.4T-A95B model card (accessed 2026-11-15)
Related NHI Mgmt Group resources: Chinese vs US Frontier Models Compared, Frontier AI Models Compared, DeepSeek Models, Zhipu Z.ai GLM Models, Moonshot AI Kimi Models and Alibaba Qwen Models.