GLM-5.3 and GLM-5.3-Flash: benchmarks, cyber capabilities, Ox Alpha, open weights and comparison with Kimi K3, Hy4, Claude and GPT-5.6 Sol

The end of August 2026 brought two very different models from the same Z.ai generation.

GLM-5.3 is the large text-only reasoning model focused on long-horizon engineering, coding agents and rapidly improving cybersecurity capabilities. Z.ai announced it on August 14, while the public weights became available on Hugging Face on August 28 after the safety evaluation and hardening period announced at launch.^[1][17]

GLM-5.3-Flash arrived later in August as a much cheaper, natively multimodal model built on a new base. It has 320B total parameters, 18B active parameters per token, a 1M-token context window, text/image/video/file input and public weights under the MIT license.^[6][8][11]

Before release, Flash was tested anonymously as:

Ox Alpha

on OpenCode and OpenRouter. Z.ai says the anonymous model quickly became the most popular model of that week and that the traffic was served on Chinese AI accelerators.^[6][7][19]

The important point is that GLM-5.3 and Flash are not simply “large” and “small” versions of the same model.

GLM-5.3 is approximately:

753B total parameters in current HF/AA metadata
~40B active parameters
text-only input
1M context
128K maximum output
reasoning always enabled
custom GLM-5.3 License

GLM-5.3-Flash is:

320B total
18B active
native multimodal input
1M context
128K maximum output
hybrid sparse + linear attention
MIT

^{[2][6][10][11]}

GLM-5.3 also shows a very large jump over GLM-5.2 in Z.ai's coding and cyber evaluations, while Flash retains much of that capability at a fraction of the API price.

The launch claims still require discipline. Most public launch tables are Z.ai-reported, agentic benchmarks depend heavily on the harness and inference budget, and competitor numbers copied into a vendor table are not automatically interchangeable with results published by those competitors under other setups.

This article therefore separates:

technical facts
vendor-reported benchmarks
independent measurements
cross-vendor comparisons

Information status: August 31, 2026.

TL;DR

Question	Verified answer
GLM-5.3 announced	August 14, 2026
GLM-5.3 public weights	available from August 28
GLM-5.3-Flash	released in late August; Artificial Analysis dates it to August 26
GLM-5.3 size	753B in current HF/AA metadata, about 40B active
GLM-5.3-Flash	320B total, 18B active
GLM-5.3 input	text
Flash input	text, image, video, file in Z.ai docs
Output	text
Context	1M for both
Maximum output	128K for both in current Z.ai docs
Reasoning	always enabled, low / high / max, default max
GLM-5.3 license	custom GLM-5.3 License
Flash license	MIT
GLM-5.3 API	$1.40 input / $0.26 cached / $4.40 output per 1M tokens
Flash list price	$0.15 / $0.03 / $0.50
Flash promo through Sep. 9, 2026	$0.075 / $0.015 / $0.25
Terminal-Bench 2.1, GLM-5.3	88.2, Z.ai-reported
Terminal-Bench 3.0	28.3, Z.ai-reported
DeepSWE v1.1	66.9, Z.ai-reported
CyberGym	84.5%, Z.ai-reported
ExploitBench	54.4%, Z.ai-reported
Ox Alpha	pre-release identity of GLM-5.3-Flash
Artificial Analysis	GLM-5.3: 60, Flash: 57 on Intelligence Index v4.1.1
Hy4 vs GLM-5.3	Tencent internal blind test: 2.99 vs 2.92
Main caveat	agent benchmarks depend on harness, tools, budget and inference configuration

What exactly happened in August?

Z.ai announced GLM-5.3 on August 14, 2026.^[1]

The unusual part was that the company did not change the base model. Z.ai states directly:

GLM-5.3 uses the same base model as GLM-5.2.
Every gain comes from post-training.

^[1][15]

The API was available at launch, but Z.ai said the weights would follow after roughly two weeks of additional safety evaluation and hardening. Hugging Face showed the planned August 28 release marker, and the weights are now publicly downloadable.^[17][5]

A few days earlier in the release window, the company had also prepared a very different model:

GLM-5.3-Flash

Flash uses a new base model, a different architecture and native multimodality.^[6][8]

GLM-5.3 is not simply a larger Flash

The models optimize for different goals.

GLM-5.3

Priority:

maximum quality in the GLM line
coding
long-horizon agents
terminal work
cyber reasoning

GLM-5.3-Flash

Priority:

cost
multimodality
inference efficiency
visual coding
office workflows
large context

Flash is not merely a distilled GLM-5.3 built from the same base. Z.ai says it was trained from a new base and uses a new architecture.^[6][8]

How many parameters does GLM-5.3 have?

This is a point where different numbers appear online.

The original GLM-5 architecture was introduced by Z.ai as:

744B total
40B active

^[16]

Current Hugging Face and Artificial Analysis metadata for GLM-5.3 report approximately:

753B total
40B active

^[5][10]

GLM-5.3 uses the same base model as GLM-5.2.^[1][15]

The 744B versus 753B difference is best treated as a model-counting/implementation metadata difference rather than evidence of a new pretraining run. In current deployment tables we use 753B / 40B, while noting that Z.ai's original architectural headline for GLM-5 was 744B / 40B.

GLM-5.3-Flash: 320B total, only 18B active

Flash is substantially smaller:

320B total
18B active
45 layers

^[6][8]

Z.ai compares it with the older GLM-4.5 family:

GLM-4.5:
355B total
32B active
92 layers

^[6]

This means almost half as many active parameters and fewer than half as many layers at a broadly similar total-parameter scale.

1M context and 128K output

Z.ai's current developer documentation gives both models:

Context Length: 1M
Maximum Output Tokens: 128K

^[2][6]

This matters for:

• large repositories,
• long agent sessions,
• multi-file analysis,
• system migrations,
• automated research,
• document-heavy workflows.

A 1M-token context window does not mean 1M tokens of perfect memory. Retrieval accuracy, instruction retention and reasoning quality at large context sizes should be tested on real workloads.

GLM-5.3 is text-only; Flash is natively multimodal

According to Z.ai's current docs, GLM-5.3 accepts:

input: text
output: text

^[2]

Flash accepts:

input:
- text
- image
- video
- file

output:
- text

^[6]

That distinction is important.

Flash is designed for workflows involving:

• screenshots,
• interfaces,
• documents,
• PDFs,
• presentations,
• spreadsheets,
• images,
• video.

Z.ai also positions it for visual coding and self-verification through rendered outputs.^[6]

Artificial Analysis currently lists text and image input in its standardized model profile, which reflects the modalities evaluated in its own system rather than overriding Z.ai's broader first-party API documentation.^[11]

Hybrid sparse + linear attention

GLM-5.3-Flash introduces a hybrid architecture combining:

sparse attention
+
linear attention

^[6][8]

Linear attention is intended to handle more local dependencies through state modeling.

Sparse attention retrieves relevant information from global context using a lightweight indexer.

The objective is lower inference cost at very long context lengths.

IndexPool

For 1M context, Z.ai introduces:

IndexPool

^[6]

The mechanism compresses four cached indexer key vectors into one through weighted pooling.

Z.ai reports that, compared with GLM-5.3, Flash reduces:

attention compute: 3.01×
KV cache: 4.44×

^[6]

These are Z.ai architectural measurements, not independently replicated infrastructure benchmarks.

Manifold-Constrained Hyper-Connections

Flash also uses:

mHC
Manifold-Constrained Hyper-Connections

^[6][8]

The technique is intended to improve scaling efficiency and information flow between layers.

For production users, the practical implication is more important than the acronym: Flash attempts to retain a large share of flagship intelligence while activating substantially fewer parameters per token.

30T-token multimodal pretraining

Z.ai says GLM-5.3-Flash was trained on its latest:

30T-token multimodal corpus

^[6][8]

This is one of the biggest differences from GLM-5.3.

GLM-5.3 keeps the GLM-5.2 base and improves primarily through post-training.

Flash received a new base, new pretraining and a different architecture.

Ox Alpha was GLM-5.3-Flash

Before its official release, Flash was tested anonymously as:

ox-alpha

on:

OpenCode
OpenRouter

^[6][19]

Z.ai says it quickly became the most popular model of the week.

That popularity claim comes from Z.ai and should not be treated as an independent global-market statistic.

The identity claim itself is clear:

Ox Alpha was the pre-release identity of GLM-5.3-Flash.

Why is the stealth test interesting?

Anonymous availability reduces some brand bias.

People selecting Ox Alpha did not necessarily know they were testing Z.ai.

But it was not a controlled scientific experiment.

Platform usage can be affected by:

• pricing,
• promotions,
• routing,
• availability,
• UI placement,
• recommendations,
• user curiosity.

The useful conclusion is therefore that Ox Alpha generated real user interest, not that anonymous usage proves it was objectively the best model.

GLM-5.3's gains came from post-training

This is one of the most important parts of the release.

Z.ai says it did not change the base model between GLM-5.2 and GLM-5.3.^[1][15]

Instead it scaled:

• environments,
• task diversity,
• post-training compute,
• long-horizon agent tasks,
• automated environment generation,
• verifiers,
• reinforcement learning.

Z.ai is developing the open-source framework:

slime

to connect training, rollout and data generation for long-horizon tasks.^[1]

“50% better coding” needs context

Z.ai markets:

+50% coding performance vs GLM-5.2

^[1][2]

That does not mean:

GLM-5.3 is 50% better on every programming task.

The claim refers to the company's private:

Z.ai Code Bench

At Max effort, Z.ai reports:

GLM-5.3: 34.5%
~75K output tokens/task

GLM-5.2: 23.4%
~96K output tokens/task

^[1]

So the vendor is claiming both greater task completion and fewer output tokens on its own benchmark.

Public coding and agent benchmarks for GLM-5.3

Z.ai publishes a broad comparison table.^[3]

Selected results:

Benchmark	GLM-5.3	GLM-5.2	Kimi K3	Opus 4.8	Fable 5	GPT-5.6 Sol
Terminal-Bench 2.1	88.2	81.0	88.3	85.0	88.0	88.8
Terminal-Bench 3.0	28.3	4.6	17.4	21.1	33.7	34.6
DeepSWE v1.1	66.9	46.2	67.5	58.0	69.7	72.7
Toolathlon Verified	73.0	59.9	76.5	76.2	74.7	74.9
AutomationBench 1.0.6	48.2	26.2	46.7	41.0	46.2	45.8
Agents' Last Exam	28.5	23.8	27.6	25.7	23.8	28.6
HLE with Tools	62.5	54.7	59.8	57.9	63.9	64.5
GDPval-AA v2	1769	1508	1682	1588	1743	1730

These are Z.ai-reported results.

Terminal-Bench 3.0 shows the largest jump

The change is:

GLM-5.2: 4.6
GLM-5.3: 28.3

^[1][3]

The numerical score is more than six times higher, but it would be incorrect to say the model is “6× more intelligent.”

The benchmark tests a particular set of terminal tasks under a specific agent harness.

Z.ai's methodology includes:

Claude Code 2.1.207
reasoning_effort=max
400K context
128K max output
avg@3
up to 600 agent turns
10h timeout

^[3]

This illustrates why agent benchmarks cannot be interpreted without the execution environment.

DeepSWE: 66.9 vs 46.2

Z.ai reports:

GLM-5.2: 46.2
GLM-5.3: 66.9
Kimi K3: 67.5
GPT-5.6 Sol: 72.7

^[3]

GLM-5.3 makes a very large generational jump, but in Z.ai's own table it does not beat Kimi K3 or GPT-5.6 Sol on this benchmark.

That is a good reminder that “SOTA” must always be attached to a specific evaluation and setup.

Terminal-Bench 2.1: effectively tied with Kimi K3

Z.ai reports:

GLM-5.3: 88.2
Kimi K3: 88.3
GPT-5.6 Sol: 88.8
Fable 5: 88.0

^[3]

The differences are tiny.

It would be misleading to describe any one of these models as crushing the others based on this benchmark.

Agents' Last Exam: almost tied with GPT-5.6 Sol

Z.ai reports:

GLM-5.3: 28.5
GPT-5.6 Sol: 28.6
Kimi K3: 27.6

^[3]

Again, the ordering is very close and changes across different evaluations.

CyberGym: GLM-5.3 leads GPT-5.6 Sol in Z.ai's table

Z.ai reports:

GLM-5.3: 84.5%
GPT-5.6 Sol: 83.6%
Fable 5: 83.8%
GLM-5.2: 77.2%
Kimi K3: 80.0%

^[1][3]

CyberGym begins from white-box source code and evaluates whether a model can identify and validate vulnerabilities by causing faults.

It is relevant to security reasoning, but a high CyberGym score does not directly measure the model's ability to compromise arbitrary live systems.

ExploitBench shows GLM-5.3's limit

Z.ai reports:

GLM-5.3: 54.4%
GLM-5.2: 24.4%
Kimi K3: 32.2%
Opus 4.8: 40.0%
Fable 5: 78.0%
GPT-5.6 Sol: 76.5%

^[1][3]

The improvement over GLM-5.2 is enormous.

At the same time, the strongest closed models in Z.ai's table remain far ahead.

That is much more informative than saying “GLM-5.3 is the best cyber model.”

ExploitGym widens the gap to the frontier

In time-normalized ExploitGym, Z.ai reports completed tasks at 2h / 6h budgets:

GLM-5.3: 105 / 130
Kimi K3: 36 / 70
Fable 5: 181 / 247
GPT-5.6 Sol: 216 / 293

^[3]

Z.ai normalizes time using per-model throughput, so the methodology is unusual and must be considered when interpreting the result.

The direction is still clear: GLM-5.3 improved sharply in cyber capability, but remains materially behind the strongest closed models on deeper exploitation tasks.

2,436 vulnerability findings in real projects

Z.ai says that, working with security teams in China, GLM-family models produced, after review, screening and deduplication:

2,436 findings
269 projects
1,097 critical + high

^[1][2]

The vendor's disclosure ledger says:

53 publicly disclosed
2,383 under embargo
107 critical
990 high
1,286 medium
53 low

^[1]

Z.ai says the oldest issue traced back to code introduced in 1981 and that the average discovered vulnerability age was 26.6 years.

These are Z.ai-reported operational results, not an independently audited global CVE dataset.

Why cyber capability needs separate governance

For enterprises, these capabilities can be useful for:

• secure code review,
• vulnerability triage,
• controlled bug reproduction,
• SAST augmentation,
• patch analysis,
• fuzzing support,
• threat modeling.

But stronger security capability also increases the importance of:

• sandboxing,
• tool-call logging,
• network isolation,
• approval gates,
• restricted test environments,
• credential minimization.

A high security benchmark score is not only a product feature. It is also an operational governance issue.

GLM-5.3-Flash coding benchmarks

Z.ai reports for Flash:^[7][8]

Benchmark	GLM-5.3-Flash	GLM-5.2
Terminal-Bench 2.1	84.3	81.0
DeepSWE v1.1	63.4	46.2
NL2Repo	56.3	48.9
Toolathlon Verified	78.4	59.9
AutomationBench 1.0.6	48.8	26.2
Agents' Last Exam	26.3	20.4
HLE with Tools	55.3	54.7
GDPval-AA v2	1773	1504

The correct conclusion is not:

Flash is equal to GLM-5.3.

It is:

Flash recovers a large share of high-end capability at radically lower API cost.

Z.ai Code Bench: Flash nearly matches Opus 4.8

On the private Z.ai Code Bench, the company reports at Max effort:

GLM-5.3-Flash: 29.0
Claude Opus 4.8: 29.5

^[6]

That is interesting, but because the benchmark is private and maintained by Z.ai, it is not independent proof that Flash broadly “matches Claude” across software engineering.

Independent Artificial Analysis: 60 vs 57

This provides a more independent reference point.

Artificial Analysis Intelligence Index v4.1.1 currently reports:

GLM-5.3 max: 60
GLM-5.3-Flash: 57

^[10][11][12]

The composite includes evaluations such as:

• GDPval-AA v2,
• Terminal-Bench 2.1,
• Humanity's Last Exam,
• GPQA Diamond,
• SciCode,
• CritPt,
• AA-Omniscience,
• AA-LCR.

Artificial Analysis also currently measures Z.ai's first-party endpoints at roughly:

GLM-5.3:
~66.5 output tokens/s
~1.6 s TTFT

Flash:
~45–49 output tokens/s
~1.5 s TTFT

^[10][11][20]

So “Flash” does not necessarily mean the highest output-token throughput on every provider. Its practical advantage is primarily lower active compute, lower API price and multimodal support.

API pricing: a major difference

Z.ai's pricing page on August 31, 2026 lists:^[9]

GLM-5.3

Input:        $1.40 / 1M tokens
Cached input: $0.26 / 1M
Output:       $4.40 / 1M

GLM-5.3-Flash — list price

Input:        $0.15 / 1M
Cached input: $0.03 / 1M
Output:       $0.50 / 1M

GLM-5.3-Flash — promotion through September 9

Input:        $0.075 / 1M
Cached input: $0.015 / 1M
Output:       $0.25 / 1M

The promotion ends September 9, 2026 at 24:00 UTC+8.^[9]

OpenRouter also lists both models, but provider pricing and routing can differ from Z.ai's first-party rates, so production cost comparisons should use the actual provider selected by the application.^[18][19]

At list prices, Flash is about:

9.3× cheaper on input
8.8× cheaper on output

than GLM-5.3.

“One-tenth the price” is a reasonable approximation

Z.ai describes Flash as delivering capability at around one-tenth the price of GLM-5.2.^[7][8]

At list prices:

GLM-5.2 input: $1.40
Flash input:   $0.15

GLM-5.2 output: $4.40
Flash output:   $0.50

^[9]

That is roughly a 9× difference, so “one-tenth” is a reasonable marketing approximation.

It does not guarantee a tenfold lower end-to-end task cost. Models may use different numbers of reasoning tokens, output tokens and tool calls.

GLM-5.3 vs Kimi K3

Z.ai's own table shows:

Terminal-Bench 2.1:
GLM-5.3 88.2
Kimi K3 88.3

DeepSWE:
GLM-5.3 66.9
Kimi K3 67.5

Terminal-Bench 3.0:
GLM-5.3 28.3
Kimi K3 17.4

AutomationBench:
GLM-5.3 48.2
Kimi K3 46.7

ExploitBench:
GLM-5.3 54.4
Kimi K3 32.2

^[3]

There is no single winner.

Kimi is slightly ahead in some coding evaluations, while GLM is much stronger in Terminal-Bench 3.0 and the security evaluations published by Z.ai.

GLM-5.3 vs Hy4 preview

Tencent ran a separate blind evaluation:

163 Tencent experts
203 real engineering tasks
0–4 scale

Average scores:

Hy4 preview: 2.99
Kimi K3: 2.94
GLM-5.3: 2.92

^[13][14]

From Hy4's perspective against GLM-5.3, Tencent reports:

46.8% wins
12.8% ties
40.4% losses

^[14]

That indicates a small advantage for Hy4 in Tencent's internal workload.

It should not be merged with Z.ai's tables into one universal leaderboard because the task set, evaluator population and harness are different.

GLM-5.3 vs GPT-5.6 Sol

In Z.ai's table, GPT-5.6 Sol remains stronger on several of the hardest evaluations:^[3]

Terminal-Bench 3.0:
28.3 vs 34.6

DeepSWE:
66.9 vs 72.7

ExploitBench:
54.4 vs 76.5

ExploitGym 6h:
130 vs 293

HLE with Tools:
62.5 vs 64.5

GLM-5.3 is slightly higher in Z.ai's CyberGym table:

84.5 vs 83.6

and AutomationBench:

48.2 vs 45.8

The data support a nuanced comparison, not the headline “open model beats GPT.”

What about Claude Opus 5?

Z.ai's GLM-5.3 launch table primarily includes:

Claude Opus 4.8
Claude Fable 5

not Claude Opus 5.^[3]

Therefore, those launch numbers should not be rewritten as:

GLM-5.3 > Claude Opus 5

without an identical contemporary benchmark and harness.

Cross-vendor tables from different dates can be useful context, but combining them as though they were one controlled evaluation creates false precision.

Open weights: the license difference matters

This is one of the most important distinctions.

GLM-5.3-Flash

Hugging Face marks:

MIT

^[8][11]

This is a conventional permissive open-source license.

GLM-5.3

GLM-5.3 does not use MIT.

It has a custom:

GLM-5.3 License

^[4]

The license grants broad rights to:

• use,
• copy,
• modify,
• merge,
• publish,
• distribute,
• sublicense,
• sell,
• run,
• deploy,
• fine-tune,
• create derivative works.

But it contains a special condition for Model as a Service.

If the licensee and its affiliates operate MaaS and exceed $10 billion in aggregate revenue in any consecutive 12-month period, they must pass a Z.ai security review before commercial use of the model or derivatives.^[4]

So the precise description is:

GLM-5.3 has public weights and broad commercial rights under a custom license, but it is not licensed identically to MIT.

Self-hosting

Official materials provide deployment paths for:

vLLM
SGLang
Transformers
Docker Model Runner

and Z.ai also lists additional ecosystems such as KTransformers, TokenSpeed and Unsloth.^[3][8]

The main GLM-5.3 Hugging Face repository is roughly:

756 GB
141 safetensors shards

^[5]

This is still a large infrastructure deployment.

Flash, at 320B total / 18B active, is significantly lighter in active compute, but it is not a typical single-consumer-GPU model either.

Chinese AI chips and a vendor-reported 3× serving improvement

Z.ai says real GLM-5.3-Flash traffic was served on a large cluster of Chinese AI accelerators.^[6][7]

The serving stack included techniques such as:

tensor parallelism
ReplaySSM
W8A8
hybrid INT8/FP8/BF16 cache
Layer Split
Encode–Prefill–Decode disaggregation

Z.ai reports:

3× end-to-end serving performance
vs the initial baseline on the same hardware

^[6][7]

The company also says per-token cost became comparable to mainstream NVIDIA GPU deployments.

These are vendor-reported infrastructure claims and should be independently validated before they drive hardware procurement decisions.

Which model should you choose?

Choose GLM-5.3 if:

• maximum GLM-family coding quality matters,
• you run long terminal tasks,
• you build coding agents,
• strong text reasoning is the priority,
• security analysis is a major use case,
• model quality matters more than API price.

Choose GLM-5.3-Flash if:

• cost is critical,
• you need vision,
• you analyze screenshots and GUIs,
• you process documents,
• you need video or file input,
• you want long context at much lower cost,
• the MIT license is an operational advantage.

Consider Kimi K3, Hy4, GPT or Claude if:

your own controlled benchmark produces a better:

cost per successful task

A launch leaderboard alone is not enough to select a production model.

Production POC checklist

Quality

• Test real internal tasks.
• Use the same harness for every model.
• Measure test pass rate.
• Measure task completion.
• Measure regressions.
• Measure out-of-scope changes.
• Test long sessions.
• Test recovery after errors.
• Validate tool calling.
• Validate structured output.

Reasoning

• Compare low, high and max.
• Remember reasoning cannot currently be disabled.
• Measure reasoning-token usage.
• Measure total output.
• Measure end-to-end latency.
• Do not use max automatically for every trivial task.

Long context

• Test 32K.
• Test 128K.
• Test 256K.
• Test 1M.
• Measure retrieval accuracy.
• Measure instruction loss.
• Test large repositories.
• Test cross-file dependencies.
• Measure KV-cache/concurrency behavior.
• Do not equate 1M context with perfect memory.

Flash multimodality

• Test screenshots.
• Test documents.
• Test charts.
• Test OCR-like workflows.
• Test GUI verification.
• Test video on your own data.

Security

• Run cyber workflows in a sandbox.
• Restrict network access.
• Log tool calls.
• Use approval gates.
• Separate production secrets.
• Do not give agents unnecessary credentials.
• Validate patches before merge.
• Treat the model as an assistant, not an autonomous security authority.

License and operations

• Review the GLM-5.3 License.
• Review the MaaS >$10B revenue condition.
• For Flash, verify MIT obligations.
• Review the selected API provider's data policy.
• Review processing region.
• Review retention.
• Measure cache hit rate.
• Measure cost per successful task.
• Define a fallback model.
• Monitor Flash pricing after the promotion ends.

POLPROG verdict

GLM-5.3 is one of the most interesting coding releases of August 2026, but not because it “beat every other model.”

It did not.

The more important result is that the same base model as GLM-5.2 was pushed much further through post-training scaling.^[1][15]

The jumps:

Terminal-Bench 3.0:
4.6 → 28.3

DeepSWE:
46.2 → 66.9

ExploitBench:
24.4 → 54.4

are large.^[1][3]

GLM-5.3-Flash demonstrates a different strategy:

320B total
18B active
native multimodal
1M context
MIT
very low API price

while scoring 57 on the independent Artificial Analysis Intelligence Index versus 60 for GLM-5.3 max.^[10][11][12]

For many real products, that can make Flash more interesting than the flagship.

The main conclusions are:

1. GLM-5.3 is a major post-training leap over GLM-5.2.
2. Its cyber capability is operationally significant, but the strongest closed models remain ahead on deeper exploitation in Z.ai's own table.
3. GLM-5.3-Flash is aggressively priced.
4. Ox Alpha was GLM-5.3-Flash.
5. Flash uses MIT; GLM-5.3 uses a custom license with a condition for extremely large MaaS operators.
6. No single benchmark settles Kimi K3, Hy4, GLM, Claude and GPT.
7. Production selection should optimize cost per successful task, not leaderboard position.

If choosing only one model for an initial POC:

GLM-5.3 for maximum text-only coding-agent quality within the GLM family.

GLM-5.3-Flash for the strongest cost, multimodality, long-context and permissive-license trade-off.