GLM-5.3-Flash

GALatest Flash

by Z.ai · GLM-5 family · best for low-cost multimodal coding and computer-use agents

Cost-OptimizedMultimodalCodingReasoningOpen-WeightsLong-Context
8.6
AI Panel Score
Value 9.8/10

GLM-5.3-Flash is Z.ai's low-cost model and the first natively multimodal model in the GLM-5 series, released on 2026-08-26 with MIT-licensed open weights on Hugging Face the same day. It is a new 320B-parameter mixture-of-experts with 18B active parameters and a hybrid of linear and sparse attention, built so that a 1M-token context stays cheap to serve. It reads images, video and files as well as text, and Z.ai pitches it as a default model for coding, computer-use and office-document agents. Z.ai reports that it beats GLM-5.2 across its benchmarks, nearly matches Claude Opus 4.8 on Z.ai's coding benchmark, and scores 57 on the Artificial Analysis Intelligence Index. The sentence a buyer needs: frontier-adjacent agent results with vision at $0.15 / $0.50, about a tenth of GLM-5.3's price and under plain MIT, but every figure is Z.ai's own and reasoning is always on.

Compare this modelAll GLM-5 versions

What's new

  • First natively multimodal GLM-5 model. Image, video and file input on the same chat endpoint, with vision trained into the coding loop: the model renders its own output, inspects it and revises.
  • New base model, not a trimmed flagship. Z.ai redesigned the architecture and training recipe and pre-trained on a new 30T-token multimodal corpus.
  • First GLM model with hybrid attention. Linear-attention layers handle local context and sparse-attention layers retrieve global context through a lightweight indexer. IndexPool compresses four indexer key vectors into one to cut indexer latency and memory at 1M tokens.
  • Manifold-Constrained Hyper-Connections (mHC) to improve scaling efficiency.
  • Efficiency against GLM-5.3 (Z.ai's figures): 3.0x less attention compute and a 4.4x smaller KV cache.
  • Vendor-reported results against GLM-5.2: DeepSWE v1.1 63.4 vs 46.2, AutomationBench 48.8 vs 26.2, Toolathlon Verified 78.4 vs 59.9, GDPval-AA v2 1,773 vs 1,504 Elo.
  • Priced at about a tenth of GLM-5.3: $0.15 / $0.50 against $1.40 / $4.40.
  • GLM-5.3-FlashX serves the model faster (up to 200 tokens/s per Z.ai) at $0.37 / $1.25.
  • Pre-release testing: Z.ai ran it anonymously as "ox-alpha" on OpenCode and OpenRouter and says it became the most popular model of the week there.
  • Served on Chinese AI chips: Z.ai built a dedicated SGLang-based inference engine for a domestic chip cluster and reports per-token cost comparable to mainstream NVIDIA GPUs.
  • Z.ai names no predecessor. Its earlier GLM-4.7-Flash remains on the price list as a free model.

Benchmarks

BenchmarkScoreSource
LMArena Elo1475lmarena.ai 2026-09-13T00:00:00.000Z
Terminal-Bench84.3%z.ai 2026-08-26T00:00:00.000Z
LMArena Coding Elo1525lmarena.ai 2026-09-13T00:00:00.000Z

AI Panel Review

Six personas, six verdicts — the same panel that reviews every product on TopReviewed.

Decision Maker8.5/10
“Near-flagship agent results, native vision and MIT weights at $0.15 in and $0.50 out make this the easiest Z.ai model to pilot, pending proof.”

GLM-5.3-Flash is the more attractive procurement story of Z.ai's August releases. It is plain MIT, so it avoids the security-review clause Z.ai added to GLM-5.3, it adds image, video and file input, and it costs about a tenth of the flagship. Z.ai's table has it ahead of GLM-5.2 everywhere and level with or ahead of GLM-5.3 on three agent benchmarks. The risks are familiar: every number is vendor-reported, Z.ai states no compliance certifications, and Z.ai serves the model on a newly built inference stack for Chinese AI chips. Six inference providers on Hugging Face offer alternative hosting.

Strategic Fit 9Vendor Risk 7.5Roadmap Confidence 8.5
Pros
  • Plain MIT license
  • multimodal input at a budget price
  • many hosts limit lock-in
Cons
  • Vendor-only benchmarks
  • no compliance certifications from Z.ai
  • first-party serving on a new hardware stack
Right for: Teams wanting a cheap default model for agents and document work that they can also self-host
Avoid if: You need audited compliance or independently verified scores before rollout
Domain Strategist9/10
“A 320B model priced like a lite tier and pitched as Opus 4.8-class: Z.ai is using its Flash model to reset what a budget model is expected to do.”

Z.ai positions GLM-5.3-Flash as a default for most workloads rather than a cut-down tier: a new base model, hybrid linear and sparse attention that Z.ai says needs a third of GLM-5.3's attention compute, vision wired into the coding loop, and a claimed Artificial Analysis Intelligence Index score of 57 that Z.ai's Chinese docs equate with Claude Opus 4.8. It is also Z.ai's proof point for Chinese AI chips, which Z.ai says now serve it at per-token cost comparable to mainstream NVIDIA GPUs. The field is dense: Qwen3.8-Flash matches its input price, and DeepSeek V4.1-Flash is cheaper on cache hits.

Competitive Positioning 9Differentiation 9Market Timing 9
Pros
  • New base model rather than a trimmed flagship
  • native multimodal agent features
  • strong early download traction
Cons
  • Crowded budget tier with open-weights rivals at similar prices
  • headline index score not independently confirmed
Right for: Products that need a cheap multimodal agent backend with open weights
Avoid if: Your positioning depends on a Western-origin model stack
Finance Lead9.5/10
“At $0.15 in and $0.50 out it costs a tenth of GLM-5.3, but always-on reasoning and an 80% cache discount keep it off the very bottom of the table.”

List price is $0.15 input, $0.50 output and $0.03 cached input on Z.ai, flat to 1M tokens, with cache storage free for now. That undercuts GPT-5.6 Luna ($0.20 / $1.20) and Gemini 3.5 Flash-Lite ($0.30 / $2.50), sits level with Qwen3.8-Flash ($0.15 / $0.47), and beats DeepSeek V4.1-Flash on output price even off-peak ($0.60), though DeepSeek's cache hits cost far less. Z.ai quotes $0.045 per Artificial Analysis task, at a discounted rate. Two things erode the headline: reasoning cannot be turned off, so short calls still produce thinking tokens, and there is no batch price.

Cost Efficiency 9.5Pricing Transparency 9Value per Dollar 9.5
Pros
  • One of the lowest output prices for a frontier-adjacent agent model
  • flat pricing to 1M tokens
  • free self-hosting under MIT
Cons
  • No batch tier
  • reasoning tokens on every call
  • shallower cache discount than DeepSeek or Qwen
Right for: High-volume agent, document and coding workloads measured by cost per task
Avoid if: Your traffic is mostly short prompts that a non-reasoning model would answer faster and cheaper
Domain Practitioner9/10
“Images, video and files through the same chat endpoint, 128 tools and a 1M window; build around reasoning that never switches off.”

Integration follows GLM-5.3: chat completions with image_url, video_url and file_url content blocks, function calling with up to 128 functions, streaming and context caching. Z.ai recommends temperature 1, top_p 0.95, reasoning_effort: "max", clear_thinking: false, and tool_stream: true when streaming. Two practical details: thinking cannot be disabled, and the API reference's multimodal request schema does not list response_format, although the model page advertises structured output, so test JSON output before depending on it. Self-hosting is lighter than the flagship (about 320 GB of FP8 weights), but the new glm5_next architecture needs framework versions that support it.

API Ergonomics 8.5Tool/Agent Support 9.5Reliability 8
Pros
  • Native multimodal input on the standard chat endpoint
  • strong vendor-reported tool-use results
  • lighter self-hosting than GLM-5.3
Cons
  • Reasoning always on
  • JSON output undocumented for multimodal requests
  • new architecture needs current serving stacks
Right for: Builders of browser-use, computer-use and document agents on a budget
Avoid if: You need low-latency non-reasoning calls or guaranteed schema-constrained output
Power User8.5/10
“Three times the Coding Plan quota of GLM-5.3, eyes for screenshots and screen recordings, and near-flagship coding; the wait is all the thinking.”

For daily use, GLM-5.3-Flash is the better fit on the GLM Coding Plan: Z.ai gives it three times the usable quota of GLM-5.3, and it can see. In ZCode it drives browser-use and computer-use agents that click through pages and check their own rendered output, and Z.ai's demos range from slide decks to Blender scenes and Godot game prototypes. Z.ai's figures put its coding close to GLM-5.3 (Terminal-Bench 2.1 84.3 against 88.2). The cost is latency, since every request thinks first, and the faster FlashX variant is API-only and not on the plan.

Output Quality 8.5Speed 7.5Everyday Usefulness 9
Pros
  • Multimodal input for screenshots and documents
  • generous Coding Plan quota
  • computer-use support in ZCode
Cons
  • Always-on reasoning latency
  • FlashX not included in the plan
  • trails GLM-5.3 on the hardest coding rows
Right for: Developers and analysts who work with screenshots, documents and long coding sessions
Avoid if: You want instant short answers
Skeptic7/10
“An Opus 4.8-level index score at a fraction of the price is Z.ai's claim, not a verified fact, and its charts skip the budget rivals buyers weigh.”

Every figure comes from Z.ai. The Artificial Analysis Intelligence Index score of 57 appears in Z.ai's blog and docs, and the $0.045-per-task figure uses a discounted price. Z.ai's base-model table puts GLM-5.3-Flash-Base roughly level with the larger GLM-5-Base (MMLU 88.1 vs 88.3, SimpleQA 33.5 vs 36.0), so most of the gains come from post-training and multimodal data. The comparison set omits Qwen3.8-Flash, DeepSeek V4.1-Flash and GPT-5.6 Luna. Z.ai concedes a larger KV cache than Kimi K3 and DeepSeek-V4-Flash, and its own table shows it trailing GPT-5.6 Terra and Gemini 3.7 Flash on DeepSWE and BabyVision, and Gemini 3.7 Flash on video.

Claim Accuracy 6.5Weakness Severity 7.5Hype vs Reality 7
Pros
  • Z.ai discloses weaknesses such as KV cache size
  • base-model table allows a like-for-like check
  • open weights allow replication
Cons
  • Vendor-only numbers
  • discounted cost-per-task framing
  • comparison set avoids direct budget rivals
Right for: Buyers who will benchmark it against Qwen3.8-Flash and DeepSeek V4.1-Flash themselves
Avoid if: You need third-party confirmation of the index score before committing

Strengths

  • Vendor-reported agent results at or above GLM-5.3 and Claude Opus 4.8 on Toolathlon Verified (78.4) and GDPval-AA v2 (1,773 Elo), at $0.15 / $0.50.
  • Native image, video and file input, with strong vendor-reported chart, document and computer-use scores (CharXiv 89.4, OfficeQA Pro 62.4, OSWorld 2.0 59.1).
  • Plain MIT weights with no commercial conditions.
  • Hybrid attention built for cheap 1M-token serving; FP8 weights about 320 GB.
  • Flat pricing to 1M tokens, unlike rivals that surcharge long prompts.
  • Three times the GLM-5.3 quota on the GLM Coding Plan.

Limitations

  • Every benchmark is vendor-reported, including the Artificial Analysis index claim.
  • Reasoning cannot be disabled, so short calls still pay for thinking tokens.
  • No batch pricing, no published rate limits, no compliance certifications.
  • Cache discount (80%) is shallower than DeepSeek's or Qwen's.
  • Trails GLM-5.3 on every coding row, and trails Gemini 3.7 Flash and GPT-5.6 Terra on BabyVision and DeepSWE, even on Z.ai's table.
  • No audio input, unlike Gemini 3.5 Flash-Lite.
  • Z.ai's comparison table leaves out the budget rivals it competes with directly.
  • Knowledge cutoff undisclosed.

Best use cases

  • Budget coding agents. Terminal and repository work in Claude Code, OpenCode or ZCode where GLM-5.3's extra few points are not worth ten times the price.
  • Visual front-end and game development. Turning screenshots, page sets or screen recordings into working front-ends or Godot prototypes, with the model rendering and checking its own output.
  • Office and finance documents. Building and checking PPTX, PDF, DOCX and XLSX deliverables, and reading PDFs as images, where OfficeQA Pro is its strongest vendor-reported vision result.
  • Computer-use and browser-use automation. Operating GUIs without APIs, backed by a vendor-reported OSWorld 2.0 score of 59.1.
  • Self-hosted multimodal deployments. Teams that want an MIT-licensed vision-language agent model small enough for a single high-memory GPU server.

Deep dive

The full research notes behind this review — verified against primary sources.

ArchitectureCapabilitiesBenchmark analysisSpeed & latencyPricing analysisDeployment & accessSafety & privacyEcosystem & tooling

Architecture

A 320B-parameter mixture-of-experts (321.3B in the Hugging Face checkpoint, including the vision encoder) that activates 18B parameters per token. Z.ai contrasts it with GLM-4.5 (355B total, 32B active, 92 layers): a similar total size, with nearly half the active parameters and half the depth. The config.json on Hugging Face (glm5_next) shows 45 layers: 34 linear-attention layers and 11 sparse-attention layers (every fourth layer), the first three with dense MLPs and the rest MoE with 288 routed experts plus one shared expert, 8 routed experts per token, hidden size 4,096 and one multi-token-prediction layer. The sparse layers use a lightweight indexer that selects the top 2,048 tokens, with IndexPool compressing four indexer keys into one. Hyper-connections run with four streams (mHC). A 24-layer vision encoder feeds images and video. The vocabulary has 154,880 tokens and the model supports 1,048,576 positions. Pre-training used a 30T-token multimodal corpus. Z.ai reports the lowest per-head attention compute among GLM-5.3, DeepSeek-V4-Flash and Kimi-K3, and concedes that its KV cache is still slightly larger than Kimi-K3's and DeepSeek-V4-Flash's. The main repository ships FP8 (E4M3) weights; zai-org/GLM-5.3-Flash-BF16 holds the BF16 checkpoint. Knowledge cutoff and training compute are undisclosed.

Capabilities

Scores lean on Z.ai's launch figures, which no independent evaluator had published at verification time, so they are set conservatively.

  • Coding (9.0): Terminal-Bench 2.1 84.3 (GLM-5.3 88.2, GPT-5.6 Terra 87.4), DeepSWE v1.1 63.4 (Terra 69.6, Gemini 3.7 Flash 65.3), NL2Repo 56.3 (Claude Opus 4.8 69.7). On Z.ai Code Bench v1.0 at max effort, 29.0 against Opus 4.8's 29.5. A few points below GLM-5.3 on every coding row.
  • Reasoning (8.8): HLE with tools 55.3, level with GLM-5.2 (54.7) and below GLM-5.3 (62.5) and Opus 4.8 (57.9). No GPQA or AIME figure.
  • Math (8.5): No math benchmark published. Set below the flagship.
  • Agentic / tool use (9.2): The strongest area on Z.ai's table: Toolathlon Verified 78.4 (highest in its comparison, above Opus 4.8's 76.2), GDPval-AA v2 1,773 Elo, AutomationBench 48.8 (Gemini 3.7 Flash 52.3) and OSWorld 2.0 59.1 for computer use.
  • Long context (8.5): 1M tokens on an architecture designed for cheap long-context serving. Linear-attention hybrids can lose precise recall, and Z.ai published no long-context benchmark to show otherwise.
  • Multilingual (8.0): English and Chinese are the tagged languages; no multilingual data.
  • Vision (8.5): CharXiv Reasoning with tools 89.4 (Opus 4.8 89.9), Chartography 78.0, MMVU 80.5 and MVBench 77.8 for video, Vision2Web 77.8. Trails Gemini 3.7 Flash and GPT-5.6 Terra on BabyVision (53.4 against 70.9 and 61.6).
  • Document / OCR (8.0): OfficeQA Pro 62.4, run on Treasury Bulletin PDFs without access to their embedded text, ahead of Opus 4.8 (48.9) on Z.ai's table. File input is supported directly.
  • Creative writing (7.0): No data. Z.ai stresses aesthetic judgment for slides and documents, which is layout rather than prose.
  • Instruction following (8.5): No IFEval figure; set just below the flagship.
  • Function calling (9.0): Up to 128 functions per request, and the Toolathlon result above.
  • Safety calibration (7.0): No safety evaluation or refusal data published. Computer-use and browser-use agents act on real systems, so guardrails are the deployer's job.
  • Real-time data (0): Z.ai lists no built-in search for this model.

Benchmark analysis

All scores are vendor-reported by Z.ai in its launch blog (2026-08-26). Competitor figures are as published in Z.ai's table. GLM-5.3 figures come from Z.ai's GLM-5.3 model card, run on the same harnesses.

Benchmark Score vs GLM-5.3 / GLM-5.2 vs Top Competitor Source
Terminal-Bench 2.1 (Claude Code harness) 84.3 88.2 / 81.0 GPT-5.6 Terra 87.4; Gemini 3.7 Flash 85.8; Claude Opus 4.8 85.0 Z.ai blog (vendor-reported)
DeepSWE v1.1 63.4 66.9 / 46.2 GPT-5.6 Terra 69.6; Gemini 3.7 Flash 65.3 Z.ai blog (vendor-reported)
NL2Repo 56.3 58.0 / 48.9 Claude Opus 4.8 69.7; DeepSeek-V4-Vision-Exp 57.7 Z.ai blog (vendor-reported)
Toolathlon Verified 78.4 73.0 / 59.9 Claude Opus 4.8 76.2; DeepSeek-V4-Vision-Exp 75.9 Z.ai blog (vendor-reported)
AutomationBench v1.0.6 48.8 48.2 / 26.2 Gemini 3.7 Flash 52.3; Claude Opus 4.8 41.0 Z.ai blog (vendor-reported)
Agents' Last Exam 26.3 28.5 / 20.4 GPT-5.6 Terra 28.0; DeepSeek-V4-Vision-Exp 27.3 Z.ai blog (vendor-reported)
HLE with tools (full set) 55.3 62.5 / 54.7 Claude Opus 4.8 57.9; DeepSeek-V4-Vision-Exp 55.1 Z.ai blog (vendor-reported)
GDPval-AA v2 (Elo) 1,773 1,769 / 1,504 DeepSeek-V4-Vision-Exp 1,675; Claude Opus 4.8 1,582 Z.ai blog (vendor-reported)
OfficeQA Pro 62.4 n/a DeepSeek-V4-Vision-Exp 57.9; Claude Opus 4.8 48.9 Z.ai blog (vendor-reported)
CharXiv Reasoning (with tools) 89.4 n/a Claude Opus 4.8 89.9; Gemini 3.7 Flash 88.7 Z.ai blog (vendor-reported)
Chartography (with tools) 78.0 n/a Claude Opus 4.8 75.0; GPT-5.6 Terra 68.0 Z.ai blog (vendor-reported)
BabyVision 53.4 n/a Gemini 3.7 Flash 70.9; GPT-5.6 Terra 61.6 Z.ai blog (vendor-reported)
MVBench (video) 77.8 n/a Gemini 3.7 Flash 82.2; GPT-5.6 Terra 75.0 Z.ai blog (vendor-reported)
MMVU (video) 80.5 n/a Gemini 3.7 Flash 82.3; GPT-5.6 Terra 75.8 Z.ai blog (vendor-reported)
OSWorld 2.0 (computer use) 59.1 n/a DeepSeek-V4-Vision-Exp 54.9; Claude Opus 4.8 54.8 Z.ai blog (vendor-reported)
Vision2Web 77.8 n/a Gemini 3.7 Flash 76.5; Claude Opus 4.8 76.1 Z.ai blog (vendor-reported)
Z.ai Code Bench v1.0 (private, max effort) 29.0 n/a Claude Opus 4.8 29.5 Z.ai blog (vendor-reported)
Artificial Analysis Intelligence Index v4.1.1 57 n/a Z.ai's Chinese docs say this equals Claude Opus 4.8's score Z.ai blog (vendor-reported)
MMLU (base model) 88.1 GLM-5-Base 88.3 DeepSeek-V4-Flash-Base 88.5 Z.ai blog (vendor-reported)
BBH (base model) 86.6 GLM-5-Base 87.4 DeepSeek-V4-Flash-Base 84.9 Z.ai blog (vendor-reported)
LiveCodeBench (base model) 37.6 GLM-5-Base 34.4 DeepSeek-V4-Flash-Base 29.9 Z.ai blog (vendor-reported)
SimpleQA (base model) 33.5 GLM-5-Base 36.0 DeepSeek-V4-Flash-Base 31.2 Z.ai blog (vendor-reported)

Only Terminal-Bench 2.1 fills a catalog compare field, on the same version stored for GLM-5.3 and GLM-5.2. The MMLU, BBH, LiveCodeBench and SimpleQA rows measure the pre-trained base model, not the released chat model, so those compare fields stay empty. HLE is with tools, unlike the no-tools HLE stored for GLM-5.2, so that field also stays empty. The Artificial Analysis score appears only in Z.ai's blog and docs, alongside a cost of $0.045 per task at a discounted price, and could not be confirmed on Artificial Analysis's own page, so the index field stays empty too. Z.ai's GLM-5.2 figures differ slightly between its two launch posts (Agents' Last Exam 20.4 here against 23.8 in the GLM-5.3 table; GDPval-AA 1,504 against 1,508). The comparison set is DeepSeek-V4-Vision-Exp, Claude Opus 4.8, GPT-5.6 Terra and Gemini 3.7 Flash; it includes none of the budget models buyers are most likely to weigh against it.

Speed & latency

No independent throughput or time-to-first-token measurement could be confirmed, so both fields are null. Z.ai's only speed figure is for GLM-5.3-FlashX, the faster serving option, at up to 200 tokens per second; it publishes no figure for standard GLM-5.3-Flash. The model's design favors long-context throughput (a third of GLM-5.3's attention compute and a much smaller KV cache, per Z.ai), but reasoning runs on every request and defaults to max, so short prompts still wait for thinking; low is the fastest setting. Z.ai serves the model on a cluster of Chinese AI chips with W8A8 quantization, mixed INT8/FP8/BF16 KV-cache quantization and separate pools for multimodal encoding, prefill and decoding, and reports a 3x end-to-end serving gain over its first baseline on that hardware. Six inference providers listed on Hugging Face also serve it, so measure on the host you plan to use.

Pricing analysis

Surface Cost Notes
API input $0.15 / 1M tok Z.ai international platform; flat to 1M tokens
API output $0.50 / 1M tok Reasoning always on
Cached input $0.03 / 1M tok 80% off input; cache storage free for a limited time
Batch (in/out) n/a No batch price on Z.ai's pricing pages
Image / video / file input token-based No separate per-image price listed
GLM-5.3-FlashX $0.37 / $1.25 $0.075 cached input; up to 200 tokens/s per Z.ai; API only
GLM Coding Plan from $18 / month Three times the usable quota of GLM-5.3; FlashX not included
Free tier none Z.ai's free models are GLM-4.7-Flash and GLM-4.5-Flash
Self-host free weights MIT
Rate limits not published Not stated on the model or pricing pages

Zhipu's Chinese docs describe the price as a tenth of GLM-5.3's, and a twentieth during a limited-time discount on its China platform. Z.ai's international pricing page shows no discount.

Deployment & access

  • Z.ai API: Model codes glm-5.3-flash and glm-5.3-flashx on Z.ai's chat completions endpoint (https://api.z.ai/api/paas/v4/chat/completions). Images go in image_url content blocks (URL or Base64 data URL, several per message), with video_url and file_url for video and documents. Text parameters match GLM-5.3. Zhipu's China platform (bigmodel.cn) has served both codes since 2026-08-26.
  • Open weights: FP8 weights at zai-org/GLM-5.3-Flash and BF16 at zai-org/GLM-5.3-Flash-BF16, under MIT (copyright Z.AI Co., Ltd), with recipes for SGLang, vLLM, TokenSpeed, Transformers, KTransformers and Unsloth. At one byte per parameter the FP8 weights are about 320 GB, less than half of GLM-5.3's, and the smaller KV cache makes long-context serving lighter.
  • License: Plain MIT with no additional conditions. GLM-5.3-Flash avoids the security-review clause of the GLM-5.3 License and the separate-license requirement for model-as-a-service businesses in Qwen3.8-Flash-Next's Qwen Community License.
  • Managed hosts: Hugging Face lists Together AI, Fireworks, DeepInfra, Novita, Baseten and Featherless AI as live inference providers.
  • Data residency: Z.ai's API data processing addendum says customer data is generally processed in Singapore.

Safety & privacy

  • Framework: Z.ai publishes no named safety framework, capability tiers or refusal metrics for GLM-5.3-Flash. Unlike GLM-5.3, whose weights Z.ai held back two weeks for safety evaluation, the Flash weights shipped on launch day.
  • Agent risk: Z.ai promotes browser-use and computer-use agents that click, type and operate desktop apps. Sandboxing, permissions and human confirmation for consequential actions are the deployer's responsibility.
  • API data handling: Z.ai's API is operated by Jingsheng Hengxing Technology Pte. Ltd. in Singapore. Its data processing addendum says Z.ai acts as a data processor on the customer's instructions, does not store the content customers input or generate (it is processed in real time), and generally processes customer data in Singapore.
  • Compliance: No SOC 2, ISO 27001, HIPAA or GDPR certification is stated.
  • Self-hosting: Under MIT, moderation, refusal behavior and logging are entirely the deployer's responsibility.

Ecosystem & tooling

Z.ai's official SDKs are zai-sdk for Python and Java, and the chat completions endpoint follows the OpenAI format. GLM-5.3-Flash is on the GLM Coding Plan for Claude Code, OpenCode and other coding agents, and in ZCode it powers Browser Use and Computer Use agents. For self-hosting, SGLang, vLLM, TokenSpeed, Transformers, KTransformers and Unsloth have GLM-5.3-Flash recipes. Together AI, Fireworks, DeepInfra, Novita, Baseten and Featherless AI serve it as live inference providers on Hugging Face. Adoption started fast: Z.ai says the anonymous "ox-alpha" test was the most popular model of the week on OpenCode and OpenRouter, and the main Hugging Face repository logged about 3.3 million downloads in its first month, more than three times GLM-5.3's.

Buyer questions

How much does GLM-5.3-Flash cost?

$0.15 per 1M input tokens, $0.50 per 1M output tokens and $0.03 per 1M cached input tokens on Z.ai, flat up to the 1M-token context. Cache storage is free for a limited time and there is no published batch discount.

What is GLM-5.3-FlashX?

A faster serving option of the same model, which Z.ai says reaches up to 200 tokens per second. It costs $0.37 / $1.25 ($0.075 cached input), about 2.5 times the standard price, and is not included in the GLM Coding Plan.

What inputs does it accept?

Text, images, video and files, sent as image_url, video_url and file_url content blocks in a chat completions request. Output is text only, with a 1M-token context and up to 128K output tokens.

How does it compare with GLM-5.3?

On Z.ai's figures it is a few points behind on coding (Terminal-Bench 2.1 84.3 vs 88.2, DeepSWE 63.4 vs 66.9) and on HLE with tools (55.3 vs 62.5), and level or ahead on Toolathlon, AutomationBench and GDPval-AA. It costs about a tenth as much, accepts images, video and files, and ships under MIT rather than the GLM-5.3 License.

Is the license different from GLM-5.3's?

Yes. GLM-5.3-Flash is plain MIT with no extra conditions. GLM-5.3's custom license adds a Z.ai security review for model-as-a-service groups with more than US$10 billion in revenue.

Can I turn off reasoning?

No. thinking.type accepts only enabled; use reasoning_effort: "low" for the lightest thinking. The default is max.

Can I self-host it?

Yes. FP8 and BF16 weights are on Hugging Face with recipes for SGLang, vLLM, TokenSpeed, Transformers, KTransformers and Unsloth. The FP8 weights are about 320 GB, so a single high-memory multi-GPU server can hold them.

Where is my data processed on Z.ai's API?

Z.ai's API data processing addendum says customer data is generally processed in Singapore and that Z.ai does not store the content customers input or generate. Z.ai lists no SOC 2, ISO 27001 or HIPAA certification.

Comparable models

Qwen3.8-Flash — Alibaba

The same $0.15 input price and a slightly lower $0.47 output price internationally, with image and video input, a 1M context, 131,072-token output and built-in web search on Model Studio. Its open weights ship as Qwen3.8-Flash-Next under the Qwen Community License, which requires a separate Qwen license before any model-as-a-service or AI work-assistant business uses them commercially. Qwen wins on output price and search; GLM-5.3-Flash wins on MIT terms and file input. Neither vendor benchmarks against the other.

DeepSeek V4.1-Flash — DeepSeek

MIT open weights with image input, a 1M context and 384K max output, at $0.30 / $1.20 peak and $0.15 / $0.60 off-peak, with cache hits at $0.006 peak and $0.003 off-peak. It keeps a non-thinking mode that GLM-5.3-Flash lacks. GLM-5.3-Flash is cheaper on output at all hours and adds video input; DeepSeek wins on cache pricing, output length and off-peak input. No shared benchmark table.

GPT-5.6 Luna — OpenAI

$0.20 / $1.20 with $0.02 cached input, a 1.05M context, 128K output, text and image input, and reasoning effort from none to max. Luna wins on governance, a no-reasoning setting and a deeper cache discount, but prompts over 272K tokens cost 2x input and 1.5x output. GLM-5.3-Flash costs 25% less on input and 58% less on output, adds video, and ships open weights. Z.ai's table compares against GPT-5.6 Terra, not Luna.

Gemini 3.5 Flash-Lite — Google

$0.30 / $2.50 with a free tier, a 1,048,576-token input limit, and text, image, video, audio and PDF input. Gemini wins on audio input, the free tier and Google's governance; GLM-5.3-Flash costs half as much on input and a fifth on output, with open weights. Z.ai's table compares against Gemini 3.7 Flash rather than Flash-Lite.

Sources

Primary references used to verify this review.

Model specs

Input price
$0.15 / Mtok
Output price
$0.50 / Mtok
Cached input
$0.03 / Mtok
Batch (in/out)
—
Context window
1M tokens
Max output
131K tokens
Knowledge cutoff
Undisclosed
Released
2026-08-26
Modalities
text, image, video → text
Output speed
Not profiled
License
Open weights (MIT)
Clouds
First-party API

Does not train on API inputs by default

Other GLM-5 versions

Last verified 2026-09-21