What it is
GLM-5.3-Flash is a separate open-weight model from Z.ai (formerly Zhipu AI), released on 26 August 2026 according to Z.ai's release notes, eight days after the larger GLM-5.3. It is a mixture-of-experts model with 320 billion parameters, of which 18 billion are used for each token, across 45 layers. Unlike GLM-5.3, it starts from a newly trained base model, and Z.ai calls it the first natively multimodal model in the GLM-5 series: it reads text, images, video and files and writes text.
The API offers a 1M-token context and up to 128K output tokens. Thinking is always on; reasoning effort can be set to low, high or max, and defaults to max. A new mix of sparse and linear attention cuts attention compute by 3.01 times and the KV cache (the memory used to hold a long conversation) by 4.44 times compared with GLM-5.3, according to Z.ai.
The weights use the plain MIT licence. On Z.ai's API it costs $0.15 per million input tokens, $0.03 for cached input and $0.50 for output. A faster GLM-5.3-FlashX, rated at 200 tokens per second, costs $0.37/$1.25, and a legal-tuned FlashX-Legal is offered only to selected enterprise customers.
Our review
GLM-5.3-Flash is the open model that makes GLM-5.3 hard to justify for most workloads. Z.ai reports 63.4 on DeepSWE v1.1 against 66.9 for the full GLM-5.3 and 46.2 for GLM-5.2, and 48.8 against 26.2 for GLM-5.2 on AutomationBench. On its in-house Z.ai Code Bench it scored 29.0 at max effort against 29.5 for Claude Opus 4.8. Those are vendor figures, and the headline comparison chart on the model card is an image rather than a table.
Independent numbers are less flattering but still strong. Artificial Analysis scores it 42 on its Intelligence Index, sixth in its open-weight size class, at $0.28 per index task. Z.ai's docs quote 57 on index version 4.1.1 at a discounted $0.045 per task, so check which version a comparison uses. Kimi K3 scores 44 on the same Artificial Analysis page set at $2.00 per task, roughly seven times the cost for two points.
The price gap with its sibling is large. At $0.15/$0.50 it is about a tenth of GLM-5.3's $1.40/$4.40, cached input is $0.03 against $0.26, and cache storage is free for now. Flash also reads images, video and files, which GLM-5.3 cannot, and plain MIT removes the custom terms on GLM-5.3's licence.
Its costs hide in the output. Thinking cannot be switched off and defaults to max effort, and Artificial Analysis measured 180M output tokens to run its index against a 140M median, at 67 tokens a second on Z.ai's API. A classification call that a non-reasoning model answers in ten tokens can return hundreds here, so the low per-token rate flatters it on short, simple jobs. Self-hosting is easier than with 2.8-trillion-parameter rivals, though about 330 GB of FP8 weights still outgrows a four-GPU 80 GB box before any KV cache.
Where it earns its place
A platform engineer running a fleet of coding agents in OpenCode or Claude Code
Cut the monthly model bill without dropping to a weak model
Point the agent harness at glm-5.3-flash on Z.ai or a host such as Together or Fireworks, both listed at the same $0.15/$0.50. A month of 2B input tokens at 80% cache hits plus 200M output tokens costs about $208, against about $1,856 for the same traffic on GLM-5.3.
A two-person startup building screenshot and PDF question-answering
Read UI captures and scanned documents in one model
Send images as image_url blocks alongside text in the same chat call; files and video are accepted too. Pass reasoning_effort low for simple lookups, because the default is max and every reasoning token is billed as output.
An engineering manager on Z.ai's GLM Coding Plan
Stretch a fixed subscription across a team
Flash draws 3x GLM-5.3's quota on the points-based Coding Plan, and off-peak and weekend calls cost half the usual points. FlashX is not on the plan yet, so the 200 tokens-a-second tier means pay-as-you-go at $0.37/$1.25.
A bank's ML team that cannot send code to an outside API
Run a capable coding model on its own hardware under a clean licence
Serve the FP8 checkpoint with vLLM or SGLang using Z.ai's published recipes; MIT means no MAU or revenue thresholds to track. Budget for an eight-GPU class server, since the weights alone are around 330 GB.
Before you commit
- Setup
- Call glm-5.3-flash on Z.ai's OpenAI-style API, or download the FP8 weights and serve them with vLLM, SGLang, TokenSpeed, KTransformers or Transformers. Z.ai recommends temperature 1, top_p 0.95 and max effort.
- Integrations
- Streaming, function calling, structured JSON output and context caching on Z.ai's API; built-in web search costs $0.01 a call. About 30 provider endpoints serve it on OpenRouter, most at $0.15/$0.50 and some FP4 hosts lower.
- Data and privacy
- Z.ai's API data terms say prompts and outputs are processed in real time and not stored, and customer data is generally processed in Singapore. The terms do not mention training either way, and no SOC 2 or ISO certification is published. Z.ai serves the model on Chinese-made accelerators.
- Cost at scale
- On Z.ai, 2B input tokens (80% cached) and 200M output tokens a month cost about $208; GLM-5.3 would be about $1,856. Self-hosting means roughly 330 GB of FP8 weights plus KV cache, which is more than four 80 GB GPUs can hold, so plan on an eight-GPU node.
- Limits
- 1M-token context and 128K output on the API; the model card does not state a context length. Thinking cannot be disabled, and Artificial Analysis measured 67 tokens a second and a 3.15-second first token on Z.ai. API rate limits are not published on the model page.
What's good
- Plain MIT licence, no revenue or attribution clauses
- DeepSWE 63.4 vs 66.9 for GLM-5.3, at about a tenth of the price
- Reads images, video and files; GLM-5.3 is text-only
- 1M-token context and 128K output on the API
What's not
- Thinking always on and verbose: 180M tokens on Artificial Analysis's run vs a 140M median
- About 330 GB of FP8 weights, so self-hosting needs a multi-GPU server
- Headline benchmarks are Z.ai's own, partly published only as an image
Best for
Teams running high-volume coding or multimodal agents who want an open model with no licence conditions and very low API costs.
Alternatives
GLM-5.3-Flash: common questions
- Is GLM-5.3-Flash worth it?
- GLM-5.3-Flash gets most of GLM-5.3's coding ability for a tenth of the token price, and its MIT licence leaves nothing for a lawyer to read. We rate it 4.3 out of 5.
- How much does GLM-5.3-Flash cost?
- Open weights under MIT. Z.ai API: $0.15 input / $0.50 output per 1M tokens ($0.03 cached); FlashX $0.37 / $1.25.
- What are the downsides of GLM-5.3-Flash?
- Thinking always on and verbose: 180M tokens on Artificial Analysis's run vs a 140M median; about 330 GB of FP8 weights, so self-hosting needs a multi-GPU server; headline benchmarks are Z.ai's own, partly published only as an image.
- Who is GLM-5.3-Flash best for?
- Teams running high-volume coding or multimodal agents who want an open model with no licence conditions and very low API costs.
- How much does GLM-5.3-Flash cost for a team or in production?
- On Z.ai, 2B input tokens (80% cached) and 200M output tokens a month cost about $208; GLM-5.3 would be about $1,856. Self-hosting means roughly 330 GB of FP8 weights plus KV cache, which is more than four 80 GB GPUs can hold, so plan on an eight-GPU node.
- Is GLM-5.3-Flash safe for company data?
- Z.ai's API data terms say prompts and outputs are processed in real time and not stored, and customer data is generally processed in Singapore. The terms do not mention training either way, and no SOC 2 or ISO certification is published. Z.ai serves the model on Chinese-made accelerators.
- What are the best GLM-5.3-Flash alternatives?
- The closest alternatives we have reviewed are GLM-5.3, DeepSeek-V4.1-Flash, Qwen3.5-397B-A17B and Kimi K3.
Answers checked Oct 23, 2026.
Make GLM-5.3-Flash? Show this rating on your site
