BigModel is Zhipu's open platform for calling GLM text, vision, image, video and audio models over an API, with knowledge bases, web search, fine-tuning and dedicated instances alongside a separate coding subscription.
Used this tool? Rate it
Used this tool? Rate it
BigModel is the developer platform of Beijing Zhipu Huazhang Technology Co., Ltd. at bigmodel.cn, where GLM text, vision, image, video and audio models are called over an API rather than used through a chat product. It positions itself as a one-stop model-as-a-service paradigm offering a feature-rich, flexible and cost-effective large-model API.
Four capabilities sit behind that phrase: calling models, deploying instances on dedicated resources, creating fine-tuning jobs from your own data, and scoring output against datasets and metrics. That span separates it from a plain inference endpoint.
Independent reporting places the operator's founding in 2019 at Tsinghua University Science Park and records a Hong Kong H-share listing in January 2026. One consequence covered below matters more: the same legal entity appears in a United States export-control record.
A recommended set of nine core models spans text generation, multimodal understanding, image generation, video generation, OCR, speech recognition, speech synthesis and vector retrieval. Text, vision, image, video and audio models share one API surface, so a single integration can reach more than one modality.
Limits are published per model rather than platform-wide. GLM-OCR accepts a single image up to 10 MB or a PDF up to 50 MB, with a ceiling of 100 pages per document. The flagship pair both carry a 1M context window with 128K maximum output.
The knowledge base stores text, images, audio and video together and answers natural-language questions across those modalities, importing from local files, URLs, Feishu documents or question-answer pairs. The version follows the embedding model and the data: omni-modal for mixed media, text-only for documents, and a QA version recognised automatically from question-answer imports.
Search is three services rather than one toggle: a raw Web Search API, search folded into chat answers, and a Search Agent that decomposes a query. Engines mix the vendor's own with Sogou and Quark, and results can be constrained by domain and time range.
One constraint breaks designs quietly. Function calling, knowledge-base retrieval and web search are mutually exclusive within one request, and when more than one is supplied only the highest-priority option takes effect, in that order.
Fine-tuning offers SFT for instruction following and DPO for preference alignment, through either LoRA or full-parameter training. Data is accepted only as JSON, one record per line, and the guidance asks for at least several hundred human-reviewed examples — that threshold, not the per-token price, sets the real cost.
The Batch API runs at half the standard rate, sits outside the normal concurrency budget, and accepts a single file of up to 50,000 requests and 100MB.
Documented paths are a REST HTTP API, first-party Python and Java SDKs, OpenAI SDK compatibility and LangChain. Migration from OpenAI code is documented as changing the API key and base URL, but the same page warns that interface differences remain in some scenarios and requires SDK 1.0.0 or newer — a real porting cost, not boilerplate.
General calls go to a versioned base URL under open.bigmodel.cn with a bearer API key, and JWT is documented where tighter security is needed. An independent third-party API profile disclaiming any vendor affiliation lists two base URLs, an international one and the China endpoint, which it labels legacy while vendor documentation still presents it as current.
Bulk offline enrichment. Classification, extraction and scoring over an existing corpus map onto the Batch API, where the trade is explicit: half price and freedom from the concurrency budget, in exchange for a completion window measured in hours.
Grounded answering over private documents. Knowledge base plus retrieval covers internal manuals and mixed-media archives; pick the omni-modal version when audio or video sits alongside text.
Real-time assistants. Streaming is the documented fit where first-token latency is visible to the user.
Prototyping at zero cost. Several Flash-class text, vision, image and video models are listed at zero for both input and output, though free models still consume concurrency.
Good fit: teams building product features on Chinese-language models. The API surface, SDKs and OpenAI compatibility make integration conventional.
Good fit: enterprises needing isolation or on-premises delivery. Dedicated instances and annual private packages are published products rather than bespoke deals.
Mixed fit: developers who mainly want a coding assistant. Subscription credits are valid only inside the vendor's approved tools, and self-built applications, websites, bots and SaaS products are explicitly directed back to the metered API. Needing both means paying into two systems.
Poor fit: teams needing vendor-side data residency outside China, since the policy commits to storing information collected in China inside China and states that cross-border transfer is not currently performed.
Poor fit: organisations that cannot clear export-control review. A January 2025 United States Commerce Department rule added the operating company and named aliases to the Entity List, requiring a licence for all items subject to the Export Administration Regulations under a presumption-of-denial review policy.
Blocked: workloads that cannot pass identity verification, which gates batch access for individuals and companies alike and grants 5 million free tokens on completion.
There is no desktop or mobile application to install; the surface is a web console plus HTTP endpoints. The console covers account setup, key management, usage statistics and billing detail, while everything operational happens over the API.
The coding subscription integrates differently, because the vendor supplies the model and someone else supplies the interface. The approved list names ZCode, Claude Code and its IDE plugin, Codex, OpenCode, Cline, Cursor, Roo, Goose, TRAE and Qoder among coding agents, plus Cherry Studio and Hermes Agent as general-purpose agents. Subscription traffic uses three base URLs distinct from the general API, one for the Anthropic message protocol and two for OpenAI-shaped protocols, and a wrong endpoint means the quota simply will not apply.
Two billing systems run side by side, and picking the wrong one is the most expensive mistake available here. Metered usage is billed per call; the coding subscription is a flat periodic fee with its own allowance, endpoints and tool restrictions.
Billing is per token, with an official guide of roughly 1:1.6 between GLM tokens and Chinese characters, and cost is assembled from uncached input, cache hits, output and cache storage. Search results count as input when search is enabled.
| Tier | Context | Input (CNY / M tokens) | Output (CNY / M tokens) | Cache hit |
|---|---|---|---|---|
| GLM-5.3 | 1M | 8 | 28 | 2 |
| GLM-5.3-Flash | 1M | 0.4 (list 0.8) | 1.4 (list 2.8) | 0.115 (list 0.23) |
| GLM-5.1 (input under 32K) | banded | 6 | 24 | 1.3 |
| GLM-5.1 (input 32K and above) | banded | 8 | 28 | 2 |
| GLM-4.5-Air (input under 32K) | banded | 0.8 | 2 | 0.16 |
| GLM-4.7-Flash | 200K | free | free | free |
Several text models are priced in bands rather than at a flat rate, so the same model costs more once input crosses 32K, and GLM-4.7 adds a further split by output length. Cache storage is marked free for a limited period, and the pricing page states it does not currently show the standard rate that will apply afterwards — an open cost that cannot be modelled today.
Image, video and speech models leave the token model entirely, metered per call, per ten-thousand characters or per minute. Search tools bill per call, and the knowledge base includes 1 GB of storage before capacity is charged hourly per gigabyte.
The subscription meters a two-layer allowance, one ceiling per five hours and another per week, across Lite, Pro and Max, with every tier reaching the current flagship pair. When it runs out the platform stops rather than spilling over: it will not draw down other resource packs or the account balance, and the allowance returns at the next five-hour boundary.
Deduction follows a fixed order, taking eligible resource packs before cash balance and, among packs of the same scope, the one expiring soonest. At the top of the range pricing turns structural: dedicated instances bill by compute unit per day, and on-premises deployment is quoted only as an order of magnitude.
The comparison that matters is lane against lane, not platform against platform. Independent developer write-ups place this subscription alongside comparable coding plans from other Chinese model vendors — a useful signal that the real comparison set is other coding subscriptions rather than general-purpose API platforms.
Three questions separate the options. Does your work live inside an approved editor or in your own application — the first is subscription territory, the second explicitly is not. How long are your typical inputs, since banded pricing charges long-context work differently on the same model. And how much porting cost can you absorb, given the caveat attached to OpenAI compatibility.
The user agreement forbids using the platform's models, prompts or generated content to develop, train, annotate, fine-tune or iterate any model outside the platform, and separately forbids improving algorithms that compete with the vendor. It also prohibits altering, removing or covering the explicit and implicit markers embedded in AI-generated content, including metadata identifiers and digital watermarks, and prohibits reverse engineering. For anyone weighing distillation or synthetic data for a house model, this is a contractual block rather than a grey area.
Prepaid balance is effectively non-refundable: the platform states it does not currently support refunds in any form, treats special cases as a reviewed exception, and excludes both consumed and already-invoiced amounts. A subscription counts as confirmed once purchased and is not refundable even if unused, and cancellation must be filed at least three days before the next charge date. Both policies argue for starting small.
The user agreement states that all data you upload, store and process through the platform belongs to you, and that apart from carrying out your service requests the platform will not make any unauthorised use or disclosure of it, the exceptions being lawful requests from state authorities and separate agreement with you. For knowledge bases, fine-tuning corpora and production prompts, that ownership-and-non-use pairing is the clause worth reading before any data is moved.
Collection covers account identifiers, device and network data, and the text, images, audio and documents submitted to models, while identity verification additionally requires a legal name, an ID number and photographs of the document. Personal information collected and generated through operations inside China is stored inside China on public cloud infrastructure, and the policy states that it is not currently transferred or stored across borders.
Two statements in the privacy policy have to be read together: personal information is not combined with algorithms for targeted push or precision marketing, yet data that has been anonymised or de-identified may be used for machine learning and model training without separate consent.
Yes. Several Flash-class models across text, vision, image and video are listed at zero for both input and output, so functional evaluation costs nothing, and identity verification grants 5 million free tokens. Free models still consume concurrency, so they test capability rather than throughput.
No. Subscription credits work only inside the approved tool list, and self-built applications, websites, bots and SaaS products are directed to the metered API instead. The two use different base URLs, so pointing your own application at the subscription endpoint does not draw on the plan.
Largely. The documented migration changes the API key and the base URL, with SDK 1.0.0 or newer required. The same page carries a vendor warning that interface differences remain in some scenarios, so budget for testing rather than assuming a drop-in swap.
Function calling, knowledge-base retrieval and web search are mutually exclusive within a single request. If more than one is supplied, only the highest-priority option takes effect in that order, so multi-tool behaviour has to be sequenced across turns.
Work stops until the next five-hour window. The platform does not fall back to other resource packs or the cash balance, which prevents unexpected charges but can halt a long task partway.
