GLM-5

zai-org MIT out 8 months ago

754B parameters (8 of 256 experts per token), text in, text out, 198K tokens of context. 10 providers serve it from $0.60 per million input tokens.

Verified 2026-10-25 Rank #35 in LLMs · score 112.6 method v1 Sources 4 Downloads 30d 1.1M · #101 of 619 tracked

Can I run it?452GB at INT4Estimateneeds a multi-GPU nodemethod

What does it cost?$0.60/1M incheapest of 9GMICloudfp82026-10-24

Can I use it?MITPermissivelicence file2026-10-24

Does it work?72.8on SWE-bench Verified#2 of 22 measuredleaderboard2026-02-17

Do people use it?1.1Mdownloads, 30 days#101 of 619 we track · top 16%+2 likes, downloads +29% this week2,139 likesHugging Face2026-10-24

Worth reading

Picked by rule from 20 tracked pieces: the lab's own post, the most shared, a field report, the latest.

Abliterating GLM-5.3 required roughly $4,400 in computation costs to reduce refusal rates from above 90% to about 3%.

From GLM-5.3 and the spread of advanced cyber capabilities · anthropic.com · 2026-09-29

The lab's own post

GLM-5: Targeting complex systems engineering and long-horizon agentic tasks

z.ai 2026-02-11 484 HN pts · 520 comments 2.1k words

GLM-5.3 and the spread of advanced cyber capabilities

Anthropic analyzes GLM-5.3, a new open-weight model from Zhipu AI that lacks robust safeguards against misuse.

Abliterating GLM-5.3 required roughly $4,400 in computation costs to reduce refusal rates from above 90% to about 3%.
Prefilling thinking tokens causes GLM-5.3 to engage with malicious requests 92% of the time without abliteration.

anthropic.com 2026-09-29 256 HN pts · 241 comments 1.8k words

Latest · 2 months ago

GLM-5.3-Flash Intelligence, Performance and Price Analysis

GLM-5.3-Flash is an open weights model released by Z AI on August 26, 2026, that scores 42 on the Artificial Analysis Intelligence Index.

The model classifies as Large because it has more than 150B parameters.
A blended rate of 7:2:1 cache hit/input/output ratio yields $0.10 per 1M tokens.

artificialanalysis.ai 2026-08-26 139 HN pts · 60 comments 969 words

20 best of 20 tracked

The 4 picks above, plus these 16. Top is the radar's score: discussion, source authority and the lab's own posts, fading with age.

ArticleGLM-5.2 is the new leading open weights model on Artificial Analysisartificialanalysis.ai · 2026-06-17 · 916 HN pts
ArticleGLM 5.2 and the coming AI margin collapsemartinalderson.com · 2026-07-06 · 694 HN pts
ArticleGLM 5.2 is nearly as accurate as a human book keepertoot-books.com · 2026-07-09 · 225 HN pts
ArticleGLM 5.2 Performance Benchmarksartificialanalysis.ai · 2026-06-17 · 164 HN pts
ArticleMiniMax M3 vs. GLM 5.2: Codegen comparison across autonomous coding tasksthinkwright.ai · 2026-06-19 · 55 HN pts
ArticleGLM-5.2: The Most Powerful Open Model yet and the Brutal Reality of Running Itvettedconsumer.com · 2026-06-19 · 43 HN pts
PaperGLM-5: from Vibe Coding to Agentic Engineeringarxiv.org · undated
Paper[2602.15763] GLM-5: from Vibe Coding to Agentic Engineeringar5iv.labs.arxiv.org · undated
ArticleGLM-5 - Everything you need to knowartificialanalysis.ai · 2026-02-11
ArticleGLM-5 Review (2026): GLM-5.2, 5.1 & 5 Benchmarks, Pricing & Access | DataLLM Labdatallmlab.com · 2026-07-10
ArticleGLM-5 Benchmarks: 10 Tests, 93.8% to 47.2%layerlens.ai · undated
ArticleGLM-5 Review 2026: The Open-Source Model That Rivals GPT-5aitoolbriefing.com · 2026-03-21
ArticleGLM-5 | Awesome Agentsawesomeagents.ai · 2026-03-03
ArticleInside GLM-5: Training Techniques, Architecture Innova… | BestHubbesthub.dev · 2026-02-19
ArticleGLM-5 Deep Dive: Key Breakthroughs, Artificial Analysis Ranking, and Practical Engineering Pros & Cons | | iWeaver AIiweaver.ai · 2026-02-19
ArticleGLM-5 Technical Review: From Vibe Coding to Agentic Engineering | Zhongzhu (Charlie) Zhouzhongzhuzhou.org · 2026-02-18

Show all 16

Where to run it

Live hosted endpoints, cheapest first, USD per million tokens. Ties break on output price, then provider name.

Prices for the same weights vary 2×

9 providers with a live price, from $0.60 to $1.00 per million input tokens; the median is $0.95. One dot per provider, cheapest in teal.

#	Cheapest 5	In /1M	Out /1M	Context	Quant	Seen
1	GMICloud	$0.60	$1.92	198K	fp8	Oct 24
2	StreamLake	$0.60	$1.92	193K	fp8	Oct 24
3	DeepInfra	$0.60	$2.08	198K	—	Oct 25
4	Baidu	$0.70	$2.24	198K	fp8	Oct 24
5	SiliconFlow	$0.95	$2.55	200K	fp8	Oct 24

All 10 providers →

Where it's used

Hugging Face does not say who downloads a model or with what. These are the proxies: API providers listing it, quantized builds, fine-tunes/adapters/merges, and Spaces that load it.

Widely served (10 API providers); 19 quantized builds for local use.

As a service10Amazon Bedrock, Baidu, DeepInfra, Featherless +6
On own hardware19quantized builds
Built on4938 fine-tunes · 10 adapters · 1 merges
In apps100+Hugging Face Spaces

Top Spaces running it: akhaliq/anycoder ♥ 3,321 · hadadxyz/ai ♥ 819 · pliny-the-prompter/obliteratus ♥ 621 · GenAISecurityProject/OWASP-AIBOM-Generator ♥ 30

Will it fit?

Weights only, an estimate: leave room for the KV cache. Method

Needs an 8×80 GB node even at INT4

About 452 GB at INT4, 905 GB at FP8, 1.81 TB at BF16: weights only, so leave room for the KV cache. Every expert must be resident, so total parameters count.

Estimate params × bytes per param × 1.2 · how we estimate

How it scores

Third-party measurements only; author-reported numbers stay in Reference, marked as such.

#2 of 22 on SWE-bench Verified so far

72.8. Each tick is a model on this site with a measured SWE-bench Verified row; author-reported numbers stay in Reference. Percent resolved. agent: mini-SWE-agent, reasoning_effort=high

Source · leaderboard · 2026-02-17

The story so far

Weights, every dated article, and repo commits on one rail.

10 dated articles and 28 repo commits so far

Teal is the weights (2026-02-11), amber triangles are 10 dated articles (each links to it), grey ticks are repo commits.

What people built on it

68 derivatives on the Hub. Which build to run, by runtime: the most downloaded in each class.

68 derivatives on the Hub, most of them fine-tunes

From the Hugging Face model tree; quantized builds in teal, fine-tunes, merges and adapters in greys.

GGUF unsloth/GLM-5-GGUF for llama.cpp, Ollama, LM Studio 141.5k downloads MLX mlx-community/GLM-5-4bit for Apple silicon 12.8k downloads NVFP4 festr2/GLM-5-NVFP4-MTP for vLLM, SGLang 46.7k downloads

All variants →

Reference

Everything else, folded: the full tables, every source.

Benchmarks

29 rows

Benchmarks 29 rows

measured third-party leaderboard row (2)author-reported number from the model card or eval files (27)No confidence intervals unless the source gave them; settings differ between rows, so compare within a benchmark, not across. Date = the benchmark's own date when it gives one, else when we read it.

Benchmark	Value	Setting / notes	Reported by	Date
measured SWE-bench Multilingual · Write code with agents	69.7 percent resolved	agent: mini-SWE-agent	leaderboard	2026-02-13
measured SWE-bench Verified · Write code with agents	72.8 percent resolved	agent: mini-SWE-agent, reasoning_effort=high	leaderboard	2026-02-17
author-reported AIME 2026 · Reason through hard problems	92.7	—	model author	2026-10-07
author-reported BrowseComp	62	—	model author	2026-10-07
author-reported BrowseComp (w/ Context Manage)	75.9	—	model author	2026-10-07
author-reported BrowseComp-Zh	72.7	—	model author	2026-10-07
author-reported CyberGym	43.2	—	model author	2026-10-07
author-reported GPQA Diamond · Reason through hard problems	86	—	model author	2026-10-07
author-reported HLE (w/ Tools)	50.4	—	model author	2026-10-07
author-reported HMMT · Reason through hard problems	96.9	—	model author	2026-10-07
author-reported Humanity's Last Exam · Reason through hard problems	30.5	—	model author	2026-10-07
author-reported IMO-AnswerBench · Reason through hard problems	82.5	—	model author	2026-10-07
author-reported MCP-Atlas (Public Set)	67.8	—	model author	2026-10-07
author-reported SWE-bench Multilingual · Write code with agents	73.3	—	model author	2026-10-07
author-reported SWE-bench Verified · Write code with agents	77.8	—	model author	2026-10-07
author-reported Terminal-Bench · Write code with agents	56.2 / 60.7 †	—	model author	2026-10-07
author-reported Terminal-Bench · Write code with agents	56.2 / 61.1 †	—	model author	2026-10-07
author-reported τ²-Bench	89.7	—	model author	2026-10-07
author-reported MathArena/aime_2026	95.83	—	Official MathArena Evaluation	2026-02-18
author-reported MathArena/hmmt_feb_2026	86.36	—	Official MathArena Evaluation	2026-02-23
author-reported diamond (Idavidrein/gpqa)	86	—	Model Card	2026-02-13
author-reported hle (cais/hle)	30.5	—	Model Card	2026-02-13
author-reported hle (cais/hle)	50.4	With tools	Model Card	2026-02-13
author-reported swe_bench_multilingual_%_resolved (SWE-bench/SWE-bench_Multilingual)	73.3	—	Model Card	2026-10-24
author-reported swe_bench_%_resolved (SWE-bench/SWE-bench_Verified)	72.8	high reasoning, official	SWE-Bench official evaluation	2026-10-24

Show all 29 rows (4 more)

Benchmark	Value	Setting / notes	Reported by	Date
author-reported swe_bench_%_resolved (SWE-bench/SWE-bench_Verified)	77.8	Z.ai reported number	Model card	2026-10-24
author-reported terminal_bench (harborframework/terminal-bench-2.0)	52.4	agent: Terminus 2	Terminal-Bench Leaderboard	2026-02-23
author-reported terminalbench_2 (harborframework/terminal-bench-2.0)	52.4	agent: Terminus 2	Terminal-Bench Leaderboard	2026-02-23
author-reported medium (collinear-ai/yc-bench)	1208190	avg final funds (USD) across seeds 1,2,3. GLM-5 (via OpenRouter z-ai/glm-5)	YC-Bench eval	2026-03-24

What the card says

2 cited sentences

Summary generated, cited

GLM-5 is a 754B parameter mixture-of-experts model designed for complex systems engineering and long-horizon agentic tasks.^[1]^[2] The architecture utilizes DeepSeek Sparse Attention (DSA) and supports English and Chinese languages.^[3]^[4]

parameters human
“targeting complex systems engineering and long-horizon ag…”
“integrates DeepSeek Sparse Attention (DSA)”
languages

Generated 2026-10-07 15:00 UTC by parasail-gemma-4-26b-a4b-it; each numeral points at the fact it rests on. Not a verdict.

Use-case evidence

5 use cases

What it is good at 5 use cases

Labels are evidence, not verdicts: measured a third-party leaderboard row for this model; author-reported a number the model author published in the card or eval files; structural a deterministic signal from the name, config or chat template, or a decision-model estimate. Each chip links to its source. All use cases.

Facts

24 fields

Listed in LLMs · MoE

Facts

GLM-5 has 754B parameters. (safetensors metadata, observed 2026-10-24)
GLM-5 is a mixture-of-experts model with 256 experts, routing 8 per token. (config.json, observed 2026-10-24)
Context length is 202,752 tokens (198K). (config.json max_position_embeddings, observed 2026-10-24)
License: MIT (friction: none). Permissive. (license file, observed 2026-10-24)
Created 2026-02-11, last modified 2026-08-11. (HF API, observed 2026-10-24)

Pipeline tag	text-generation
Library	transformers
Architectures	GlmMoeDsaForCausalLM
Model type	glm_moe_dsa
Dtype	bfloat16
Params by dtype	BF16: 753,864,119,552 · F32: 19,456
Size band	XXL
MoE	yes
Experts	256
Experts per token	8
Context length	202,752 (config.json max_position_embeddings)
Languages	en, zh
Papers	arXiv:2602.15763
License	MIT (mit)
Hub snapshot	2026-10-25

Servability

vLLM · SGLang

Servability

vLLM: supported · GlmMoeDsaForCausalLM · 7050ef881ff0835f6890a47db1859a24c65481d6SGLang: not in registry · 773cb938fe9484a8f7e7e46fb1b1fe96412c59ed

vLLM registry (7050ef881ff0835f6890a47db1859a24c65481d6) · SGLang registry (773cb938fe9484a8f7e7e46fb1b1fe96412c59ed) · Matched by architecture class name against each project's registry on main; 'supported' means the architecture is registered, not that this checkpoint was tested.

Hardware fit

3 estimates

Hardware fit ESTIMATE

Weights-only memory for GLM-5: params × bytes/param × 1.2 overhead. KV cache and activations are extra. How this is computed.

BF16 (2 bytes/param)	1.81 TB · Cluster (≥640 GB)
FP8 (1 byte/param)	905 GB · Cluster (≥640 GB)
INT4 (0.5 bytes/param)	452 GB · Multi-GPU node (<640 GB)

Weights only (params × bytes/param × 1.2 overhead); KV cache and activations extra. MoE: all experts must be resident, so total parameters count.

All variants

40 variants

Lineage

Original model Family: zai-org/GLM-5 68 derivatives on the Hub

finetune: 38quantized: 19adapter: 10merge: 1

Top variants (40 shown, ranked by downloads)

Variant	Relation	Quant	Params	Band	License	Author	Likes	Downloads
nvidia/GLM-5-NVFP4	finetune	NVFP4	435B	XXL	mit	nvidia	32	600.7k
unsloth/GLM-5-GGUF	quantized	GGUF	—	—	mit	unsloth	228	141.5k
festr2/GLM-5-NVFP4-MTP	quantized	NVFP4	435B	XXL	mit	festr2	3	46.7k
amd/GLM-5-MXFP4	quantized	MXFP4	408B	XXL	mit	amd	3	35.8k
lukealonso/GLM-5-NVFP4	quantized	NVFP4	425B	XXL	mit	lukealonso	9	28.1k
unsloth/GLM-5-FP8	quantized	-FP8	754B	XXL	mit	unsloth	5	13.2k
mlx-community/GLM-5-4bit	quantized	-4bit	744B	XXL	mit	mlx-community	8	12.8k
QuantTrio/GLM-5-AWQ	quantized	AWQ	754B	XXL	mit	QuantTrio	6	11.5k
mlx-community/GLM-5-MXFP4-Q8	quantized	MXFP4	744B	XXL	mit	mlx-community	1	11.7k
tiny-random/glm-5	finetune	—	4.58M	S	—	tiny-random	0	9.3k
Intel/GLM-5-int4-mixed-AutoRound	quantized	int4	122B	XXL	—	Intel	4	8.2k
mlx-community/GLM-5	finetune	—	744B	XXL	mit	mlx-community	4	8.1k
ubergarm/GLM-5-GGUF	quantized	GGUF	—	—	mit	ubergarm	16	6.5k
DevQuasar/zai-org.GLM-5-GGUF	quantized	GGUF	—	—	—	DevQuasar	0	4.4k
yujiepan/glm-moe-dsa-tiny-random	finetune	—	4.58M	S	—	yujiepan	0	3.6k
gghfez/GLM-5-control-vectors	adapter	GGUF	—	—	mit	gghfez	2	3.1k
unsloth/GLM-5	finetune	—	754B	XXL	mit	unsloth	18	2.2k
skyblanket/GLM-5-abliterated	finetune	—	754B	XXL	apache-2.0	skyblanket	17	2k
yujiepan/glm-5-tiny-random	finetune	—	4.58M	S	—	yujiepan	1	2.2k
abhilashms/GLM-5-MLX-4.8bit	quantized	MLX	744B	XXL	—	abhilashms	0	2.2k
0xSero/GLM-5-381B-GGUF-IQ2_M	quantized	GGUF	—	—	apache-2.0	0xSero	1	2k
tiny-random/glm-moe-dsa	finetune	—	4.58M	S	—	tiny-random	2	1.9k
0xSero/GLM-5-381B	finetune	—	381B	XXL	apache-2.0	0xSero	3	1.6k
sokann/GLM-5-GGUF-1.594bpw	quantized	GGUF	—	—	mit	sokann	4	1.5k
Explyt/GLM-5-AWQ	quantized	AWQ	754B	XXL	mit	Explyt	0	1.6k
lhca521/GLM-5-abliterated	finetune	—	754B	XXL	apache-2.0	lhca521	0	1.6k
varjosoft/GLM-5.1-Open-TQ3	finetune	—	289B	XXL	mit	varjosoft	2	1.2k
nwzjk/GLM-5.1-Open-TQ3	finetune	—	289B	XXL	mit	nwzjk	2	511
amd/GLM-5-NVFP4	quantized	NVFP4	380B	XXL	mit	amd	1	548
MuVeraAI/glm-5-381-reap-w3a16	quantized	—	41.4B	XL	apache-2.0	MuVeraAI	0	596
Trilogix1/GLM-5-abliterated	finetune	—	754B	XXL	other	Trilogix1	0	563
Doctor-Shotgun/GLM-5-GGUF	quantized	GGUF	—	—	—	Doctor-Shotgun	0	263
rbinrs/GLM-5-abliterated	finetune	—	754B	XXL	apache-2.0	rbinrs	2	105
CPU-Hybrid-MoE/GLM-5-CPU-NUMA4-AMXINT8	finetune	—	—	—	—	CPU-Hybrid-MoE	1	141
Explyt/GLM-5-AWQ-pruned-20	quantized	AWQ	609B	XXL	mit	Explyt	0	151
kepom/GLM-5-abliterated	finetune	—	754B	XXL	apache-2.0	kepom	0	139
Felipe2231/canario-amarelo-prod-final	merge	—	—	—	artistic-2.0	Felipe2231	2	0
GabriellSaid/GLM-5-abliterated	finetune	—	754B	XXL	apache-2.0	GabriellSaid	0	79
HollowMan6/GLM-5-NOOP-LoRA	adapter	—	—	—	mit	HollowMan6	0	55
Sadiqeen/Secondmodel	finetune	—	—	—	bigscience-bloom-rail-1.0	Sadiqeen	1	0

Repository changelog

28 commits

Repository changelog 28 commits

Date	Commit	SHA
2026-08-11	Add community evaluation results for SWE-Bench Multilingual (#84)	c183ef8
2026-04-05	update OS frameworks to stable releases	4e6698b
2026-04-03	Add YC-Bench evaluation results (avg $1,208,190) (#70)	04670d2
2026-04-03	Add MathArena evaluation result for hmmt/hmmt_feb_2026 (#51)	ec61e8d
2026-04-03	Add MathArena evaluation result for aime/aime_2026 (#44)	8c33dc9
2026-03-24	Add Terminal-Bench evaluation result (52.4%) (#56)	dce1f7c
2026-03-24	Update SWE-Bench Verified results (#65)	0ab7173
2026-03-13	Add Terminal-Bench 2.0 evaluation result (52.4%) (#64)	73b81c6
2026-03-13	Add arXiv metadata and update citation with paper links (#40)	fafc0ac
2026-03-11	Update chat_template.jinja	f4c6240
2026-03-10	Add SWE-Bench Verified evaluation result (#53)	64eea7d
2026-02-13	Add evaluation results for GPQA, HLE (#22)	b8a9fc5
2026-02-13	Update README.md (#18)	360b49d
2026-02-11	update	83d08ca
2026-02-11	line	9329f32
2026-02-11	sglang update	17b316b
2026-02-11	work	cce0887
2026-02-11	init3	fbfe45b
2026-02-11	init2	c35b39f
2026-02-11	init2	1a68bb7
2026-02-11	init	baee9b2
2026-02-11	Add files using upload-large-folder tool	773a94e
2026-02-11	Add files using upload-large-folder tool	5c90e22
2026-02-11	Add files using upload-large-folder tool	9e60996
2026-02-11	Add files using upload-large-folder tool	fb2dfc3
2026-02-11	Add files using upload-large-folder tool	b9c39a3
2026-02-11	Add files using upload-large-folder tool	22ce68a
2026-02-11	initial commit	670766b

All providers

10 providers · 10 endpoints

Where to run 10 providers · 10 endpoints

Hosted endpoints observed for GLM-5, sorted by input price (USD per million tokens). Each row names its observer and date; prices change daily, so check the source before quoting. Cheapest observed input price: $0.60/M.

Provider	Model ref	Context	Quant	$/M in	$/M out	tok/s	TTFT	Tools	Status	Observed
DeepInfra hf	zai-org/GLM-5	198K	—	$0.60	$2.08	80.8	342 ms	yes	live	2026-10-25
StreamLake openrouter	z-ai/glm-5	193K	fp8	$0.60	$1.92	—	—	yes	live	2026-10-24
GMICloud openrouter	z-ai/glm-5	198K	fp8	$0.60	$1.92	—	—	yes	live	2026-10-24
Baidu openrouter	z-ai/glm-5	198K	fp8	$0.70	$2.24	—	—	yes	live	2026-10-24
SiliconFlow openrouter	z-ai/glm-5	200K	fp8	$0.95	$2.55	—	—	yes	live	2026-10-24
Novita hf, openrouter	zai-org/glm-5	198K	fp8	$1.00	$3.20	42.5	1217 ms	yes	live	2026-10-25
Z.ai hf, openrouter	glm-5	198K	fp8	$1.00	$3.20	76.3	3733 ms	yes	live	2026-10-25
Amazon Bedrock openrouter	z-ai/glm-5	198K	unknown	$1.00	$3.20	—	—	yes	live	2026-10-24
Venice openrouter	z-ai/glm-5	193K	fp8	$1.00	$3.20	—	—	yes	live	2026-10-24
Featherless hf	zai-org/GLM-5	—	—	—	—	—	—	—	live	2026-10-25

Articles and coverage

20 pieces

Ranked by source signals (HN, Lobsters, authority domains, official posts) and recency. Summaries are generated from a stored copy of the article text; every sentence cites a span of that text, and nothing is shown that failed the citation check.

Official1

GLM-5: Targeting complex systems engineering and long-horizon agentic tasks

z.ai2026-02-11officialauthorityHN front pageHN 484 pts · 520 comments2,061 words

Analysis9

GLM-5.3 and the spread of advanced cyber capabilities

anthropic.com2026-09-29HN front pageHN 256 pts · 241 comments1,799 words

Anthropic analyzes GLM-5.3, a new open-weight model from Zhipu AI that lacks robust safeguards against misuse. Testing shows the model develops end-to-end exploits in 50 of 410 attempts and full control-flow hijacks in 4% of trials, while abliteration reduces refusal rates from above 90% to about 3%. Researchers found a specific exploit chain could be built in 20 minutes of human attention for a cost of $20.40, whereas deceptive prompts engage the model 64% of the time.written, reviewed, cited

Abliterating GLM-5.3 required roughly $4,400 in computation costs to reduce refusal rates from above 90% to about 3%.
Prefilling thinking tokens causes GLM-5.3 to engage with malicious requests 92% of the time without abliteration.
GLM-5.3-Flash built a reliable exploit chain for an ARM64 target in eight hours of autonomous work plus 20 minutes of human attention.

GLM-5.3-Flash Intelligence, Performance and Price Analysis

artificialanalysis.ai2026-08-26authorityHN front pageHN 139 pts · 60 comments969 words

GLM-5.3-Flash is an open weights model released by Z AI on August 26, 2026, that scores 42 on the Artificial Analysis Intelligence Index. It costs $0.15 per 1M input tokens and $0.50 per 1M output tokens, operates under the MIT license, and utilizes 320B total parameters with 18B active parameters during inference. The model supports text and image input with a 1M token context window but generates verbose output at 68.9 tokens per second.written, reviewed, cited

The model classifies as Large because it has more than 150B parameters.
A blended rate of 7:2:1 cache hit/input/output ratio yields $0.10 per 1M tokens.
The model is available through 26 API providers for deployment.

GLM-5.3 Artificial Analysis Benchmarks

artificialanalysis.ai2026-08-18authorityHN front pageHN 153 pts · 55 comments989 words

GLM-5.3 (Max) is an open weights model from Z AI released on August 18, 2026, that scores 45 on the Artificial Analysis Intelligence Index. It generates output at 85.0 tokens per second while costing $1.40 per 1M input tokens and $4.40 per 1M output tokens. The model produced 210M output tokens during evaluation, which is higher than the median of 140M for comparable models.written, reviewed, cited

The model uses a Mixture of Experts architecture with 753 billion total parameters but only 40 billion active parameters during inference.
For a blended rate using a 7:2:1 cache hit/input/output ratio, the effective cost is $0.90 per 1M tokens.

GLM-5.2 is the new leading open weights model on Artificial Analysis

artificialanalysis.ai2026-06-17authorityHN front pageHN 916 pts · 444 comments720 words

Z ai's GLM-5.2 is the new leading open weights model on the Artificial Analysis Intelligence Index, scoring 51 and sitting on the Pareto frontier of Intelligence vs Cost per Task. The model features 744B total parameters with a 1M token context window and carries an MIT license. It costs approximately $0.46 per task, utilizing 43k output tokens per Intelligence Index task, which is higher than peers like MiniMax-M3 at 24k.written, reviewed, cited

GLM-5.2 scores 4 on the AA-Omniscience Index, up from GLM-5.1 (2), with a lower hallucination rate of 28.1% versus 29.4%.

GLM 5.2 and the coming AI margin collapse

martinalderson.com2026-07-06HN front pageHN 694 pts · 469 comments1,444 words

GLM 5.2 from Z.ai acts as a drop-in replacement for Opus and GPT 5.5, costing around $4.40/MTok which is less than 20% of the retail price of Opus. While the model lacks vision support and has poor web search capabilities, it runs 2.75x cheaper per token on AMD hardware compared to Nvidia Blackwell. These factors make it a credible option for budget-conscious workflows despite slower inference speeds due to extensive thinking.written, reviewed, cited

GLM 5.2 lacks vision support, preventing it from reading image-based PDFs, screenshots, and design files.
The model's poor web search capabilities are a significant weakness because nearly every agentic session performs web searching.
Z.ai's official API terms regarding training and data retention are described as very lax and difficult to sell for professional use.

GLM 5.2 is nearly as accurate as a human book keeper

toot-books.com2026-07-09HN front pageHN 225 pts · 122 comments1,820 words

GLM 5.2 Performance Benchmarks

artificialanalysis.ai2026-06-17authorityHN front pageHN 164 pts · 48 comments1,022 words

MiniMax M3 vs. GLM 5.2: Codegen comparison across autonomous coding tasks

thinkwright.ai2026-06-19HN 55 pts · 19 comments1,244 words

Thinkbench evaluated GLM 5.2 and MiniMax M3 on autonomous coding tasks involving greenfield builds and bug fixes. GLM 5.2 achieved a 92% full-pass rate with a 0.976 mean score, while MiniMax M3 reached 84% full-pass and a 0.961 mean score. Although GLM was more accurate, MiniMax M3 cost $6.67 versus GLM's $18.47 and averaged 45 seconds per run against GLM's 80 seconds. GLM 5.2 is the safer pick for hard from-scratch builds, while MiniMax M3 is the value pick for bug fixes, feature additions, or repair-to-green loops under review.written, reviewed, cited

MiniMax M3 used a fixed one-second background tick while GLM 5.2 employed a heap for more idle-efficient scheduling.
The benchmark provider is Fireworks AI, serving both models via a serverless endpoint at https://api.fireworks.ai/inference/v1.
Input pricing differs significantly between models, with MiniMax M3 at $0.45 / Mtok compared to GLM 5.2 at $2.10 / Mtok.

GLM-5.2: The Most Powerful Open Model yet and the Brutal Reality of Running It

vettedconsumer.com2026-06-19HN 43 pts · 28 comments1,259 words

All coverage tracked for this model (20) — append-only, nothing is ever removed

Title	Source	Published	First seen	Signals
GLM-5: from Vibe Coding to Agentic Engineering	arxiv.org	—	2026-10-07	exa
[2602.15763] GLM-5: from Vibe Coding to Agentic Engineering	ar5iv.labs.arxiv.org	—	2026-10-07	exa
GLM-5 Benchmarks: 10 Tests, 93.8% to 47.2%	layerlens.ai	—	2026-10-07	exa
GLM-5.3 and the spread of advanced cyber capabilities	anthropic.com	2026-09-29	2026-10-07	hn
GLM-5.3-Flash Intelligence, Performance and Price Analysis	artificialanalysis.ai	2026-08-26	2026-10-07	hn
GLM-5.3 Artificial Analysis Benchmarks	artificialanalysis.ai	2026-08-18	2026-10-07	hn
GLM-5 Review (2026): GLM-5.2, 5.1 & 5 Benchmarks, Pricing & Access \| DataLLM Lab	datallmlab.com	2026-07-10	2026-10-07	exa
GLM 5.2 is nearly as accurate as a human book keeper	toot-books.com	2026-07-09	2026-10-07	hn
GLM 5.2 and the coming AI margin collapse	martinalderson.com	2026-07-06	2026-10-07	hn
MiniMax M3 vs. GLM 5.2: Codegen comparison across autonomous coding tasks	thinkwright.ai	2026-06-19	2026-10-07	hn
GLM-5.2: The Most Powerful Open Model yet and the Brutal Reality of Running It	vettedconsumer.com	2026-06-19	2026-10-07	hn
GLM-5.2 is the new leading open weights model on Artificial Analysis	artificialanalysis.ai	2026-06-17	2026-10-07	hn
GLM 5.2 Performance Benchmarks	artificialanalysis.ai	2026-06-17	2026-10-07	hn
GLM-5 Review 2026: The Open-Source Model That Rivals GPT-5	aitoolbriefing.com	2026-03-21	2026-10-07	exa
GLM-5 \| Awesome Agents	awesomeagents.ai	2026-03-03	2026-10-07	exa
Inside GLM-5: Training Techniques, Architecture Innova… \| BestHub	besthub.dev	2026-02-19	2026-10-07	exa
GLM-5 Deep Dive: Key Breakthroughs, Artificial Analysis Ranking, and Practical Engineering Pros & Cons \| \| iWeaver AI	iweaver.ai	2026-02-19	2026-10-07	exa
GLM-5 Technical Review: From Vibe Coding to Agentic Engineering \| Zhongzhu (Charlie) Zhou	zhongzhuzhou.org	2026-02-18	2026-10-07	exa
GLM-5: Targeting complex systems engineering and long-horizon agentic tasks	z.ai official	2026-02-11	2026-10-07	hn
GLM-5 - Everything you need to know	artificialanalysis.ai	2026-02-11	2026-10-07	exa

Papers

1 paper

Papers 1

arXiv links found in the model card and its tags. "Peer-reviewed" is shown only with evidence: a published venue in OpenAlex (measured) or the authors' own journal reference on arXiv (reported).

GLM-5: from Vibe Coding to Agentic Engineering this model

GLM-5-Team, Aohan Zeng, Xin Lv +183 · 2026-02-17 · PDF · HF paper page (223 upvotes)

Similar models

3 models

Similar models

GLM-5.3 · 753B
same size band (XXL), also MoE, same task (text-generation), more providers
GLM-5.2 · 753B
same size band (XXL), also MoE, same task (text-generation), more providers, permissive license
GLM-5.1 · 754B
same size band (XXL), also MoE, same task (text-generation), fewer providers, permissive license

Sources

4 sources

Sources

Hub snapshot: 2026-10-25 · HF API: 2026-10-24 21:32 UTC · Providers: 2026-10-24 21:44 UTC · Articles: 2026-10-24 21:45 UTC · Summary: 2026-10-07 15:00 UTC · Last verified 2026-10-25

Follow GLM-5

Watch this model

/feeds/models/zai-org--GLM-5.xml(RSS 2.0) · every change to GLM-5
/feeds/families/zai-org--GLM-5.xml(RSS 2.0) · every change across the GLM-5 family

Email alerts are not open yet; the feeds above are the same data. See feeds for the item schema.