GLM-5.3-Flash
The first natively multimodal model in the GLM-5 series. With 320B total parameters and just 18B active parameters, it outperforms GLM-5.2 across benchmarks and real-world workloads at one-tenth the price, while approaching Claude Opus 4.8 on coding and agentic benchmarks.
Z.ai’s own description of the model, from GLM-5.3-Flash model card. Served here as zai/glm-5.3-flash.
On this roster: Pick it for fast agentic loops at the lowest price here: big-model reasoning, tools and vision on a 262K window.
Speculative decoding on this model is drafted by DFlash 2 (inco.ai), used in production with the authors' written permission. The speculative-decoding drafter on Hugging Face.
What Z.ai built it for
in the vendor’s words- For the first time in the GLM series, we introduce a hybrid architecture combining sparse and linear attention, sharply reducing long-context serving costs while preserving precise long-context capabilities.
- The model also adopts Manifold-Constrained Hyper-Connections (mHC) to further improve scaling efficiency. Together with our latest 30T-token multimodal pre-training corpus, these changes enable GLM-5.3-Flash to deliver more intelligence with less compute.
Quoted from GLM-5.3-Flash model card, retrieved 1 Sept 2026. Reasoning is always on, with a three-level effort dial (low, high, max) that defaults to max. The card's own instruction for benchmark and leaderboard reproduction is to keep the default max, which is also what an omitting request gets here.
Published benchmarks
Z.ai’s figures, not ours| Benchmark | What it measures | Score | How Z.ai ran it |
|---|---|---|---|
| Terminal Bench 2.1 | Agentic terminal coding | 84.3 | The card's footnote: evaluated in Claude Code 2.1.207 with temperature=1.0, top_p=1, max_new_tokens=65536 with 6h timeout. Mode: not stated for this row. |
| DeepSWE v1.1 | Agentic coding, real-world software engineering | 63.4 | The card's footnote: run using the mini-swe-agent harness with temperature=0.95, top_p=1.0, timeout=6h and 400K context. Mode: not stated for this row. |
| GDPVal-AA v2 | Real-world economically valuable tasks | 1773 | The card's footnote: models are evaluated by Artificial Analysis. Mode: not stated for this row. |
| HLE w/ Tools | Reasoning with tools | 55.3 | The card's footnote (full set): temperature=1.0 and top_p=0.95, maximum generation length 163,840 tokens, maximum context length 300,000 tokens with a context management strategy, GPT-5.6-luna (medium) as judge. Mode: not stated for this row. |
| AutomationBench | Workflow automation | 48.8 | The card's footnote: AutomationBench v1.0.6, incorporating the fix for the null-type handling issue introduced in PR #13. Mode: not stated for this row. |
| Agents' Last Exam | Agentic reasoning | 26.3 | as published (the card's footnote for this row is blank). Mode: not stated for this row. |
Every score above is Z.ai’s own published result for this model, copied from GLM-5.3-Flash model card on 1 Sept 2026. We did not run these benchmarks and we do not restate them as our own. Each vendor ran its own harness with its own judge, so these tables are not a league table: scores from two different cards are not comparable, even where the benchmark name matches. Each model here shows only its own maker’s table, for that reason.
Recommended usage
what Z.ai recommendsSampling
For recommended inference settings - the default path, and what we serve, Z.ai recommends temperature=1.0, top_p=0.95. That is what this endpoint sends when your request omits them, so the default request shape is the vendor’s recommendation rather than ours. An explicit value in your request always wins.
Thinking
Reasoning is always on and streams before the answer; there is no off switch on this family. The reasoning_effort parameter takes low, high, or max and defaults to max - Z.ai's own instruction for reproducing its benchmark scores is to keep the default, and that is also what an omitting request gets here. Lower efforts spend fewer thinking tokens; reasoning tokens bill as output either way.
Context length
Z.ai gives the native window as 1,048,576 tokens (the checkpoint's trained window). We serve a pinned 262,144-token window of it, and the window we publish is the window we serve: input and output share it, and an oversized prompt is refused with a clean 429 before it primes rather than failing mid-request. The full trained window on this architecture needs a larger placement than this deployment runs; serving it is engineering in progress, not a checkbox. Reasoning is mandatory on this model and defaults to maximum effort, and reasoning tokens bill as output - send a generous max_tokens and stream.
What to point it at
Z.ai names these workloads for this model:
- “it outperforms GLM-5.2 across benchmarks and real-world workloads at one-tenth the price, while approaching Claude Opus 4.8 on coding and agentic benchmarks”
- “sharply reducing long-context serving costs while preserving precise long-context capabilities”
Recommendations on this page are Z.ai’s, from GLM-5.3-Flash model card, retrieved 1 Sept 2026. We default to them rather than to a house preference; where our default would differ, this page says so.
What it does on this endpoint
our measurements · not vendor figuresThese figures are ours, measured on the live endpoint: observations, not service guarantees, and not to be confused with the vendor scores above. The protocol, the vantage, the build each figure is pinned to and how to check it against the system_fingerprint in your own response are all on the measurements page. Run your own and use those instead.
Price
per million tokens · no request fee · no subscription| Input | Cached input | Output |
|---|---|---|
Limited-time launch promotion: 50% off the struck-through list price, matching the model vendor's own promotion.
One price, no tiers. The cached rate applies only to tokens your responses report as cached_tokens — see the prompt caching card below for where that stands on this model. Full price sheet and worked examples.
Capabilities
zai/glm-5.3-flashThree API formats, one base URL
OpenAI Chat Completions, OpenAI Responses, and Anthropic Messages, all served from https://api.tiyuvta.ai/v1. Point an existing client at it and set the model id; there is no product-specific SDK.
Tools
tools and tool_choice work as on every model here, and the tool-call round trip is part of the release gate on every deploy. response_format is not available on this model: its reasoning channel cannot be disabled, so the API refuses structured-output requests with a clear error instead of returning malformed JSON.
Vision
Images go in as image_url content parts and bill as ordinary input tokens; vision costs nothing extra.
Prompt caching
Prompt caching is on for this model. When a request's prompt begins with a prefix the serving box still holds from one of your earlier requests (a multi-turn conversation, a repeated system prompt, a re-sent document), those tokens are restored instead of re-read and billed at the cached-input rate; the response reports them as cached_tokens. In our 8-turn measurement on the serving box the cache engaged every turn: time to first token stayed flat at 1.2 to 1.6 s while the same conversation without the cache rose to 4.4 s by turn 8, with 4,839 of 5,429 prompt tokens cached on that turn. Residency is bounded by a per-box budget and shared with other traffic, so after a long pause or on a busy box a turn can come back cold and bill at the normal input rate. You are only ever billed the cached rate for tokens your response actually reports as cached. Measured on the serving box under a six-way mixed load: warm strict-prefix hit rate 0.747 (68 of 91 warm turns) at the shipped 8 GiB cache budget, eviction-bound; a cache-read price is only honest with its hit rate beside it.
Exactness-gated serving
Speculative decoding is verified byte-identical to plain decode on the serving configuration before a build takes traffic; speed never changes the output. The correctness gates.
Where it fits best
use it for- Long-context agent loops: the 262,144-token window we serve is the window we publish, and admission refuses an oversized prompt with a 429 before it primes rather than failing mid-request
- Fast agentic turns: speculative decoding is drafted by DFlash 2 (inco.ai) under written permission from its authors - it is what makes the decode fast
- Reasoning-heavy work where you want to see the reasoning: this model always reasons, the channel is exposed, and reasoning tokens are billed as output - send a generous max_tokens and stream
- Big-model behaviour at a mid-size bill: 320B total parameters, ~18B active per token, on our own 4-bit mint of the official checkpoint
Also on the same key and balance: Ornith-1.5 35B-A3B, Qwen3.8 27B, Qwen3-Embedding-8B, Qwen3-Reranker-8B, Step-3.7-Flash.
First request
zai/glm-5.3-flashcurl https://api.tiyuvta.ai/v1/chat/completions \
-H "Authorization: Bearer $TIYUVTA_KEY" \
-H "Content-Type: application/json" \
-d '{"model": "zai/glm-5.3-flash",
"messages": [{"role": "user", "content": "Hello"}]}'The default request reasons before it answers, at Z.ai’s recommended sampling: best quality for hard problems.
- Weights and vendor sources
- huggingface.co/tiyuvta/GLM-5.3-Flash-NVFP4, MIT. Z.ai trained this model; we serve it at its full native context in quantized precision. Benchmark scores and usage recommendations on this page come from GLM-5.3-Flash model card, retrieved 1 Sept 2026.
- Machine-readable metadata
GET https://api.tiyuvta.ai/v1/modelsfor supported parameters, and the provider feed atGET https://api.tiyuvta.ai/models?schema=openrouter.- Rate limits
- Per account, for abuse control, raised on request; capacity scales with demand rather than sitting behind a published ceiling.