GLM 5.3 Flash
Z.ai · released Aug 25, 2026 · zai-org/GLM-5.3-Flash
Input: text, images and video. Output: text.InputOutput
- Type
- Open weightsMIT License
- Params
- 321B
- Context
- 1.3M
about 983K words of context
Our take
Written Sep 26, 2026GLM 5.3 Flash is a downloadable model you can run yourself, and it is at its best when the job is agentic: picking the right tool and finishing the task. It is at its worst when the job is holding to a format under instruction, and we list no licence for it, so the terms need checking at the source.
Who should pick it
Reach for it on agent work where the model has to choose the right tool and see the task through, and on maths-heavy or coding prompts, where it places well inside the top tenth of a large field. Long documents need not be split up first. Skip it if the job depends on holding to a format or a constraint, if a task needs recovery after a command fails, or if you need licence terms confirmed before you build on it.
The case for it
- 9th of 52 on Arena Agent · Tool use as of 15 Sep 2026, a board that scores whether the model calls the right tool and does not invent one, so this is evidence for tool selection rather than for the wider task.
- 10th of 52 on Arena Agent · Task outcome as of 15 Sep 2026, which scores finishing the job the session set out to do rather than the step in front of it.
- 5th of 153 on Arena Maths as of 13 Sep 2026 and 14th of 158 on Arena Coding as of 13 Sep 2026, so maths-heavy and coding prompts are where to point it first.
- The request capacity takes a long document or a stack of documents alongside the question, though recall across all of it is unverified in our data.
The case against it
- 51st of 51 on LiveBench Instruction Following as of 25 Jun 2026, a board of constrained-rewriting tasks, so format-critical work needs checking rather than trusting.
- 38th of 52 on Arena Agent · Recovery as of 15 Sep 2026, the weak end of its own agentic results, so a task that has to survive a failed command is the wrong place to start.
- The weights are downloadable but no licence is supplied, so commercial use, changes and redistribution cannot be confirmed from here and need checking at the source.
00
How good is it?
An open text model for everyday questions, coding and calling tools to carry out requests.
Good at
- getting answers to everyday questionsArena Text (overall) · 26th of 168
- writing and completing codeArena Coding · 22nd of 168
- calling tools to carry out requestsArena Agent · Tool use · 9th of 55
EverydayGeneral questions and everyday reasoning
4 of 5
Arena Text (overall)26th of 168 · 1474
Arena Hard Prompts 22nd of 168Arena Maths 10th of 163LiveBench Data Analysis 25th of 58LiveBench Reasoning 48th of 58LiveBench Mathematics 50th of 58
CodingWriting and fixing code on its own
4 of 5
Arena Coding22nd of 168 · 1523
Arena Code (WebDev) 17th of 95LiveBench Coding 22nd of 58
AgenticPlanning, calling tools, staying on task
2.5 of 5
Arena Agent27th of 55 · 0.002
LiveBench Agentic Coding 20th of 58
WritingDrafting and rewriting prose
3 of 5
Arena Creative Writing43rd of 168 · 1433
LiveBench Language 36th of 58
How it behaves in an agent loop
Tool usereaches for the right one, and does not invent one9th of 55
Steerabilitydoes what it was asked, and changes course when told27th of 55
Recoverygets back on track after a command fails40th of 55
Task outcomefinishes what the session set out to do11th of 55
Placings on Arena's agent boards, from live sessions people ran themselves. A model can lead on one of these and sit mid-field on the others.
Other boards it appears on
Arena Instruction Following 21st of 168LiveBench 44th of 58LiveBench Instruction Following 58th of 58Arena Agent · Tool use 9th of 55Arena Agent · Task outcome 11th of 55Arena Agent · Steerability 27th of 55Arena Agent · Recovery 40th of 55
Boards this model appears on that none of the ratings above are built on.
Every published score for this model20 scoresEvery figure we hold, from 20 boards, with who ran it and a link to the source — including the boards no rating above is built on.
01
Can you run it yourself?
A card many people ownToo large
GeForce RTX 4090 · 24 GB
Weights at 202.6 / 24 GBest
Usable contextNot calculated for spilled setups
Decode speedNot estimated for spilled setups
Too large for this card. The weights do not fit even with part of them offloaded to system memory.
One step upToo large
Apple M1 Pro (16-core GPU) · 32 GB
Weights at 202.6 / 32 GBest
Usable contextNot calculated for spilled setups
Decode speedNot estimated for spilled setups
Too large for this card. The weights do not fit even with part of them offloaded to system memory.
Comfortable fit
On a MacFits in memory
Apple M3 Ultra (80-core GPU) · 512 GB
Weights at 202.6 / 512 GBest
Spare memory173.5 GB spare
Usable context262K of 1.3M
Decode speed3 tok/sest
Room to spare. 173.5 GB spare means a 10% error in the size would not change the answer.
Your hardware
Checking your profile…
Memory use by level
Against a 24 GB card.
What is quantisation? →202.6 GBest
237.7 GBest
355 GBest
This model on every device we track71 devicesThe Q4 build most people download, on each device: what the weights come to, how much context the memory leaves, and whether it runs. Smallest device that runs it first.
Check against your own machine → · Where to rent it hosted →
02
Or rent it from someone else
Prices checked 3 hours ago — each listing carries its own date.
Some hosts sell this model at two prices: on their own price list (“direct”) and on their OpenRouter listing (“through OpenRouter”). Where the two differ, the row shows both, each with the date we last read it.
Cheapest published offer
Relace, through OpenRouter
Cheapest of the 9 listings we can compare like for like — at 1M of context, out of 33 in the table below. 10 cheaper rows there are outside that comparison: a different quantisation or a different context length.
- per 1M tokens
- $0.035 in / $0.50 out
- Context served
- 1M
- Throughput
- ~40 tok/s
| Provider | In / out per 1M tokens | Context | Throughput | Trains on prompts | Logs prompts | Zero retention |
|---|---|---|---|---|---|---|
| DeepInfrafp4Direct and through OpenRouter | $0.15 / $0.50directchecked 3 hours ago$0.075 / $0.25through OpenRouterchecked 3 hours ago | 1M131K max reply through OpenRouter | 18 tok/sthrough OpenRouter | DirectUnknownThrough OpenRouterNo | DirectUnknownThrough OpenRouterNo | DirectUnknownThrough OpenRouterConfirmed |
| Novita AIfp8Direct and through OpenRouter | $0.15 / $0.50directchecked 3 hours ago$0.084 / $0.28through OpenRouterchecked 3 hours ago | 1M131K max reply through OpenRouter | 17 tok/sthrough OpenRouter | DirectUnknownThrough OpenRouterNo | DirectUnknownThrough OpenRouterNo | DirectUnknownThrough OpenRouterConfirmed |
| StreamLakefp8Through OpenRouter | $0.087 / $0.29checked 3 hours ago | 1M128K max reply | 36 tok/s | No | Yesunknown period | Unknown |
| OpenInferencefp4Through OpenRouter | $0.020 / $0.30checked 3 hours ago | 1M944K max reply | 15 tok/s | No | No | Confirmed |
| GMICloudfp8Through OpenRouter | $0.090 / $0.30checked 3 hours ago | 1M944K max reply | 29 tok/s | No | Yesunknown period | Unknown |
| Near AIfp8Through OpenRouter | $0.10 / $0.35checked 3 hours ago | 1M944K max reply | 10 tok/s | No | No | Confirmed |
| Phalafp8Through OpenRouter | $0.12 / $0.40checked 3 hours ago | 1M131K max reply | 23 tok/s | No | No | Confirmed |
| Decartfp4Through OpenRouter | $0.13 / $0.42checked 3 hours ago | 1M944K max reply | 73 tok/s | No | No | Confirmed |
| Io Netfp8Through OpenRouter | $0.14 / $0.45checked 3 hours ago | 262K66K max reply | 31 tok/s | No | No | Confirmed |
| Inceptronfp8Through OpenRouter | $0.23 / $0.45checked 3 hours ago | 1M944K max reply | 15 tok/s | No | No | Confirmed |
| Modalnvfp4Through OpenRouter | $0.15 / $0.50checked 3 hours ago | 1M944K max reply | 84 tok/s | No | No | Confirmed |
| RekaThrough OpenRouter | $0.15 / $0.50checked 3 hours ago | 262K236K max reply | 54 tok/s | No | No | Confirmed |
| RelaceThrough OpenRouter | $0.035 / $0.50checked 3 hours ago | 1M131K max reply | 40 tok/s | No | No | Confirmed |
| SiliconFlowfp8Through OpenRouter | $0.15 / $0.50checked 3 hours ago | 1M262K max reply | 35 tok/s | No | No | Confirmed |
| Crusoefp4Through OpenRouter | $0.15 / $0.50checked 3 hours ago | 1M944K max reply | 81 tok/s | No | No | Confirmed |
| DigitalOcean GradientThrough OpenRouter | $0.15 / $0.50checked 3 hours ago | 1M944K max reply | 24 tok/s | No | No | Confirmed |
| Parasailfp8Through OpenRouter | $0.15 / $0.50checked 3 hours ago | 1M944K max reply | 65 tok/s | No | No | Confirmed |
| Together AIThrough OpenRouter | $0.15 / $0.50checked 3 hours ago | 1M944K max reply | 81 tok/s | No | No | Confirmed |
| OpenRouterOpenRouter's own listing | $0.15 / $0.50checked 3 hours ago | 1M | not measured | Unknown | Unknown | Unknown |
| CoreWeavenvfp4Through OpenRouter | $0.15 / $0.50checked 3 hours ago | 1M944K max reply | 82 tok/s | No | No | Confirmed |
| Basetenfp8Through OpenRouter | $0.15 / $0.50checked 3 hours ago | 1M131K max reply | 100 tok/s | No | No | Confirmed |
| AtlasCloudfp8Through OpenRouter | $0.15 / $0.50checked 3 hours ago | 1M131K max reply | 34 tok/s | No | Yesunknown period | Unknown |
| Venice AIThrough OpenRouter | $0.15 / $0.50checked 3 hours ago | 1M131K max reply | 21 tok/s | No | No | Confirmed |
| Fireworks AIThrough OpenRouter | $0.15 / $0.50checked 3 hours ago | 1M944K max reply | 48 tok/s | No | No | Confirmed |
| FriendliThrough OpenRouter | $0.15 / $0.50checked 3 hours ago | 1M944K max reply | 77 tok/s | No | Yesunknown period | Unknown |
| Z.AIfp8Through OpenRouter | $0.15 / $0.50checked 3 hours ago | 1M131K max reply | 36 tok/s | No | No | Confirmed |
| NextBitfp8Through OpenRouter | $0.17 / $0.55checked 3 hours ago | 1M128K max reply | 31 tok/s | No | No | Confirmed |
| InferenceNetfp4Through OpenRouter | $0.050 / $0.60checked 3 hours ago | 1M262K max reply | 34 tok/s | No | No | Confirmed |
| Sail Researchusfp4Through OpenRouter | $0.045 / $0.60checked 3 hours ago | 1M131K max reply | 31 tok/s | No | No | Unknown |
| Morphfp8Through OpenRouter | $0.20 / $0.70checked 3 hours ago | 1M944K max reply | 106 tok/s | No | No | Confirmed |
| Fireworks AIusThrough OpenRouter | $0.23 / $0.75checked 3 hours ago | 1M944K max reply | 70 tok/s | No | No | Confirmed |
| WaferThrough OpenRouter | $1.00 / $0.75checked 3 hours ago | 1M944K max reply | 30 tok/s | No | No | Confirmed |
| Cloudflare Workers AIThrough OpenRouter | $0.30 / $1.00checked 3 hours ago | 1M944K max reply | 28 tok/s | No | Yesunknown period | Unknown |
DeepInfra
Novita AI
DigitalOcean Gradient
Together AI
OpenRouter
Fireworks AI
Morph
Cloudflare Workers AIAcross the 33 listings we hold: 32 say they do not train on prompts (2 of them only through OpenRouter), 0 say they do and 1 does not say. 26 appear in the zero-retention registry we check (2 of them only through OpenRouter); the rest are unknown to us.
What each host's API supports
From the parameter list each endpoint publishes. Streaming is omitted: nothing we hold reports it, for any model.
| Provider | Tool calling | JSON output | Strict schema |
|---|---|---|---|
| DeepInfrafp4Direct and through OpenRouter | ✓ | ✓ | ✓ |
| Novita AIfp8Direct and through OpenRouter | ✓ | ✓ | ✗ |
| StreamLakefp8Through OpenRouter | ✓ | ✓ | ✗ |
| OpenInferencefp4Through OpenRouter | ✓ | ✓ | ✓ |
| GMICloudfp8Through OpenRouter | ✓ | ✓ | ✗ |
| Near AIfp8Through OpenRouter | ✓ | ✓ | ✓ |
| Phalafp8Through OpenRouter | ✓ | ✓ | ✓ |
| Decartfp4Through OpenRouter | ✓ | ✓ | ✓ |
| Io Netfp8Through OpenRouter | ✓ | ✓ | ✗ |
| Inceptronfp8Through OpenRouter | ✓ | ✓ | ✓ |
| Modalnvfp4Through OpenRouter | ✓ | ✓ | ✓ |
| RekaThrough OpenRouter | ✓ | ✓ | ✓ |
| RelaceThrough OpenRouter | ✓ | ✗ | ✗ |
| SiliconFlowfp8Through OpenRouter | ✓ | ✓ | ✗ |
| Crusoefp4Through OpenRouter | ✓ | ✓ | ✓ |
| DigitalOcean GradientThrough OpenRouter | ✓ | ✓ | ✓ |
| Parasailfp8Through OpenRouter | ✓ | ✓ | ✓ |
| Together AIThrough OpenRouter | ✓ | ✓ | ✓ |
| OpenRouterOpenRouter's own listing | ✓ | ✓ | ✓ |
| CoreWeavenvfp4Through OpenRouter | ✓ | ✓ | ✓ |
| Basetenfp8Through OpenRouter | ✓ | ✓ | ✓ |
| AtlasCloudfp8Through OpenRouter | ✓ | ✗ | ✗ |
| Venice AIThrough OpenRouter | ✓ | ✓ | ✓ |
| Fireworks AIThrough OpenRouter | ✓ | ✓ | ✓ |
| FriendliThrough OpenRouter | ✓ | ✓ | ✓ |
| Z.AIfp8Through OpenRouter | ✓ | ✓ | ✗ |
| NextBitfp8Through OpenRouter | ✓ | ✓ | ✓ |
| InferenceNetfp4Through OpenRouter | ✓ | ✓ | ✓ |
| Sail Researchus · fp4Through OpenRouter | ✓ | ✓ | ✓ |
| Morphfp8Through OpenRouter | ✓ | ✓ | ✓ |
| Fireworks AIusThrough OpenRouter | ✓ | ✓ | ✓ |
| WaferThrough OpenRouter | ✓ | ✓ | ✓ |
| Cloudflare Workers AIThrough OpenRouter | ✓ | ✓ | ✓ |
Tool calling: 33 of 33 listings say yes. JSON output: 31 of 33 listings say yes, 2 say no. Strict schema: 25 of 33 listings say yes, 8 say no.
03
Models people weigh against GLM 5.3 Flash
Other sizes and versionsGLM 4.7 Flash
SucceedsGLM 4.7 Flash
04
When we formed this view
Recent changes
Oct 1, 2026Price changeGLM 5.3 Flash repriced across 2 hosts: InferenceNet input down 50%, Wafer output up 50%
What moved
GLM 5.3 Flash moved on 2 hosts: Wafer: input +33% ($0.75 → $1.00 per 1M tokens), output +50% ($0.50 → $0.75 per 1M tokens); InferenceNet: input −50% ($0.100 → $0.050 per 1M tokens), output +33% ($0.45 → $0.60 per 1M tokens), cache read +60% ($0.030 → $0.048 per 1M tokens)Sep 30, 2026Price changeGLM 5.3 Flash repriced across 5 hosts: OpenInference output down 18%, InferenceNet output up 61%
What moved
GLM 5.3 Flash moved on 5 hosts: InferenceNet: input +11% ($0.090 → $0.100 per 1M tokens), output +61% ($0.28 → $0.45 per 1M tokens), cache read +50% ($0.020 → $0.030 per 1M tokens); Inceptron: input +50% ($0.150 → $0.225 per 1M tokens), cache read +33% ($0.060 → $0.080 per 1M tokens); Relace: input +40% ($0.025 → $0.035 per 1M tokens), cache read +40% ($0.025 → $0.035 per 1M tokens); OpenInference: output −18% ($0.300 → $0.247 per 1M tokens); Morph: input −10% ($0.20 → $0.18 per 1M tokens), output −10% ($0.70 → $0.63 per 1M tokens)Sep 29, 2026Price changeGLM 5.3 Flash repriced across 4 hosts: OpenInference input down 60%, Inceptron input up 25%
What moved
GLM 5.3 Flash moved on 4 hosts: OpenInference: input −60% ($0.050 → $0.020 per 1M tokens), output −40% ($0.50 → $0.30 per 1M tokens), cache read −50% ($0.020 → $0.010 per 1M tokens); Relace: input −38% ($0.040 → $0.025 per 1M tokens), cache read +67% ($0.015 → $0.025 per 1M tokens); Morph: cache read +31% ($0.0306 → $0.0400 per 1M tokens); Inceptron: input +25% ($0.12 → $0.15 per 1M tokens), cache read +50% ($0.040 → $0.060 per 1M tokens)Sep 28, 2026Price changeGLM 5.3 Flash repriced across 4 hosts: AtlasCloud down 23% on all rates, Wafer input up 74%
What moved
GLM 5.3 Flash moved on 4 hosts: Wafer: input +74% ($0.43 → $0.75 per 1M tokens); AtlasCloud: input −23% ($0.150 → $0.116 per 1M tokens), output −23% ($0.500 → $0.385 per 1M tokens), cache read −23% ($0.030 → $0.023 per 1M tokens); Inceptron: input +9% ($0.11 → $0.12 per 1M tokens), cache read −43% ($0.070 → $0.040 per 1M tokens); Morph: input −6% ($0.16 → $0.15 per 1M tokens), output −6% ($0.568 → $0.536 per 1M tokens), cache read −6% ($0.0325 → $0.0306 per 1M tokens)Sep 27, 2026Price changeHost Relace raised GLM 5.3 Flash input pricing by 75%
What moved
input +75% ($0.040 → $0.070 per 1M tokens), cache read +33% ($0.015 → $0.020 per 1M tokens)Sep 26, 2026Price changeGLM 5.3 Flash repriced across 3 hosts: Relace input down 43%, Wafer input up 523% · machine-readable source ↗
What moved
GLM 5.3 Flash moved on 3 hosts: Wafer: input +523% ($0.069 → $0.430 per 1M tokens), output +43% ($0.35 → $0.50 per 1M tokens); Relace: input −43% ($0.070 → $0.040 per 1M tokens), output +79% ($0.28 → $0.50 per 1M tokens), cache read −25% ($0.020 → $0.015 per 1M tokens); OpenInference: cache read +33% ($0.015 → $0.020 per 1M tokens)Sep 25, 2026Price changeGLM 5.3 Flash cut across 3 hosts, by up to 50% at OpenInference (input)
What moved
GLM 5.3 Flash moved on 3 hosts: OpenInference: input −50% ($0.100 → $0.050 per 1M tokens), cache read −40% ($0.025 → $0.015 per 1M tokens); Inceptron: input −27% ($0.15 → $0.11 per 1M tokens), output −10% ($0.50 → $0.45 per 1M tokens); Wafer: input −22% ($0.089 → $0.069 per 1M tokens)Each date is the day we first saw the change, or the day the maker announced it.
What we do not know about this model yet
- We hold no measured file for it, so all 3 sizes on this page are calculated from the parameter count.
- Nothing we hold says whether an endpoint streams, so we do not show it either way.
- 1 of 33 listings does not say whether it trains on prompts, and 2 answer only through OpenRouter, not for their own listing.
- We hold no batch or off-peak rate for any of its listings.
- We hold a decode speed for it, but no prompt-processing (prefill) figure, so how long the input side of a job takes is unknown to us.
05
Licence and identifiers
What the licence allowsMIT License, what it allows commercially, and the identifiers you need to pull this model — its Hugging Face repo, our slug and a machine-readable card.
Licence
MIT License
Open, few conditionsCommercial use allowed
Fully permissive: do anything with attribution. No patent grant, unlike Apache-2.0.
Identifiers
- Hugging Face
- zai-org/GLM-5.3-Flash
- Architecture
- Dense
- Takes in, gives back
- Text, images and video in, text out
- Catalogue slug
- z-ai-glm-5-3-flash