Up to date
zai-org GLM License (glm-5.3) out 2 months ago
753B parameters (8 of 256 experts per token), text in, text out, 1M tokens of context. 38 providers serve it from $0.075 per million input tokens.
Verified 2026-10-25 Rank #6 in LLMs · score 131.9 method v1 Sources 5 Downloads 30d 2M · #62 of 619 tracked
Changed this week: Cheapest input price rose 30.2%: $0.058 → $0.075 per 1M (OpenRouter).2026-10-24 · Cheapest input price rose 37.2%: $0.058 → $0.079 per 1M (OpenRouter).2026-10-24 · 12 new quantized builds on the Hub (88 in all).2026-10-24
Why did the open-weights model cross the road?
To get to the other side of the paywall.
.-""""-.
.' ___ '.
/ .'` `'. \
| / ___ \ |
\ \_/ \_/ /
'. | | .'
'-...-'
\
O
/|\
/ \
~~~~~~~~~~~~~~~~~~~
~ ~ ~ ~ ~ ~ ~No playable frames in the reply.
The ringer: the same four playful tests for every model, replies shown as received and checked by code. A joke, not a benchmark; never ranked.2 of 4 pass · parasail-glm-53 on parasail · run 2026-10-09 · temperature 0 · cost $0.104 · 2026-10 thing: supermarket · all models →
Picked by rule from 33 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%.
Z.ai released GLM-5.3, an open-weights model that improves coding and cyber capabilities through post-training scaling.
Anthropic analyzes GLM-5.3, a new open-weight model from Zhipu AI that lacks robust safeguards against misuse.
A team ran GLM 5.3 Flash exclusively for September but failed after 2B tokens due to infrastructure limits and a prototype error.
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.
Z.ai introduces GLM-5.3-Flash, the first natively multimodal model in the GLM-5 series, scoring 57 at $0.045 per task.
The 5 picks above, plus these 20. Top is the radar's score: discussion, source authority and the lab's own posts, fading with age.
Live hosted endpoints, cheapest first, USD per million tokens. Ties break on output price, then provider name.
37 providers with a live price, from $0.075 to $2.10 per million input tokens; the median is $0.91. One dot per provider, cheapest in teal.
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 (38 API providers); heavily run locally (88 quantized builds).
Top Spaces running it: embedl/hfviewer ♥ 181 · FineEnvs/simulation-rl-environments ♥ 12 · llamameta/glm5.3-xliberated-free ♥ 4 · akhaliq/GLM-5.3 ♥ 3
Weights only, an estimate: leave room for the KV cache. Method
About 452 GB at INT4, 904 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
Claim-matched tests, graded by code. Receipts, not a score.
{
"providers": [
{
"provider": "parasail",
"endpoint_model": "parasail-glm-53",
"source_url": "https://api.parasail.io/v1/chat/completions",
"ttft_ms_median": 5119,
"ttft_any_ms_median": 574,
"tokens_per_s_median": 189.6,
"runs": [
{
"ttft_s": 4.048,
"ttft_any_s": 0.57,
"tokens_per_s": 149.7,
"latency_s": 4.93,
"tokens_out": 738,
"reasoning_chars": 2639,
"finish": "stop"
},
{
"ttft_s": 5.119,
"ttft_any_s": 0.589,
"tokens_per_s": 219.1,
"latency_s": 5.61,
"tokens_out": 1229,
"reasoning_chars": 5044,
"finish": "stop"
},
{
"ttft_s": 5.126,
"ttft_any_s": 0.574,
"tokens_per_s": 189.6,
"latency_s": 5.69,
"tokens_out": 1078,
"reasoning_chars": 4121,
"finish": "stop"
}
]
}
],
"note": "Single client, wall clock; tok/s counts every output token including hidden reasoning."
} {
"calls": [
{
"name": "get_weather",
"args": {
"city": "Lisbon",
"date": "2026-10-14"
}
},
{
"name": "check_calendar",
"args": {
"date": "2026-10-14"
}
},
{
"name": "book_table",
"args": {
"date": "2026-10-14",
"party_size": 2,
"restaurant": "Cervejaria Ramiro",
"time": "19:30"
}
}
],
"checks": {
"get_weather": true,
"check_calendar": true,
"book_table": true
},
"final": "All done! Here's a summary:\n\n- **Weather:** Clear skies in Lisbon on Wednesday, 2026-10-14, with only a 5% chance of rain and a high of 22°C — no rain expected.\n- **Calendar:** Your free evening slot is 19:30–21:30.\n- **Booking:** Table for 2 at Cervejaria Ramiro at 19:30 on 2026-10-14.\n\nYour confirmation code is **TBL-4471**. Enjoy your dinner!",
"fixture": "receipts.errand.v1"
} Weights, every dated article, and repo commits on one rail.
Teal is the weights (2026-08-25), amber triangles are 10 dated articles (each links to it), grey ticks are repo commits.
104 derivatives on the Hub. Which build to run, by runtime: the most downloaded in each class.
From the Hugging Face model tree; quantized builds in teal, fine-tunes, merges and adapters in greys.
Everything else, folded: the full tables, every source.
| Benchmark | Value | Setting / notes | Reported by | Date |
|---|---|---|---|---|
| measured Terminal-Bench · Write code with agents | 41.82 percent (accuracy, mean of trials) | agent: Claude Code, reasoning_effort=max | leaderboard | 2026-08-14 |
| author-reported Agents' Last Exam (ALE-CLI) | 28.5 | — | model author | 2026-10-07 |
| author-reported AutomationBench (v1.0.6) | 48.2 | — | model author | 2026-10-07 |
| author-reported CyberGym | 84.5 | — | model author | 2026-10-07 |
| author-reported DeepSWE · Write code with agents | 66.9 | — | model author | 2026-10-07 |
| author-reported ExploitGym (2h / 6h) | 105 / 130 | 2h / 6h | model author | 2026-10-07 |
| author-reported FrontierSWE · Write code with agents | 78.1 | — | model author | 2026-10-07 |
| author-reported GDPval-AA v2 | 1769 | — | model author | 2026-10-07 |
| author-reported HLE w/ Tools | 62.5 | — | model author | 2026-10-07 |
| author-reported NL2Repo · Write code with agents | 58 | — | model author | 2026-10-07 |
| author-reported PostTrainBench | 39.8 | — | model author | 2026-10-07 |
| author-reported ProgramBench (Almost Solved) | 19 | — | model author | 2026-10-07 |
| author-reported SWE-Marathon (v1.1) | 42.5 | — | model author | 2026-10-07 |
| author-reported Terminal-Bench · Write code with agents | 88.2 | — | model author | 2026-10-07 |
| author-reported Terminal-Bench · Write code with agents | 28.3 | — | model author | 2026-10-07 |
| author-reported Terminal-Bench · Write code with agents | 54.4 | — | model author | 2026-10-07 |
| author-reported Toolathlon · Call tools and APIs | 73 pass@1 | — | model author | 2026-10-07 |
| author-reported ale-cli-pass-rate (agents-last-exam/agents-last-exam) | 28.5 | — | zai-org/GLM-5.3 model card | 2026-08-31 |
| author-reported deep_swe (datacurve/deep-swe) | 66.9 | — | Model Card | 2026-10-24 |
| author-reported hle (cais/hle) | 62.5 | With tools. | Model Card | 2026-10-24 |
| author-reported terminalbench_2_1 (harborframework/terminal-bench-2.1) | 88.2 | harness: claude code | Model Card | 2026-10-24 |
| author-reported terminalbench_3 (harborframework/terminal-bench) | 28.3 | harness: claude code | Model Card | 2026-10-24 |
| author-reported toolathlon_verified (hkust-nlp/Toolathlon) | 73 | — | zai-org/GLM-5.3 model card | 2026-08-31 |
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.
| Pipeline tag | text-generation |
|---|---|
| Library | transformers |
| Architectures | GlmMoeDsaForCausalLM |
| Model type | glm_moe_dsa |
| Dtype | bfloat16 |
| Params by dtype | BF16: 2,103,729,152 · F32: 19,456 · F8_E4M3: 751,226,191,872 |
| Size band | XXL |
| MoE | yes |
| Experts | 256 |
| Experts per token | 8 |
| Context length | 1,048,576 (config.json max_position_embeddings) |
| Languages | en, zh |
| Papers | arXiv:2602.15763 |
| License | GLM License (glm-5.3) (other) |
| Hub snapshot | 2026-10-25 |
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.
Weights-only memory for GLM-5.3: 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) | 904 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.
quantized: 88finetune: 16
| Variant | Relation | Quant | Params | Band | License | Author | Likes | Downloads |
|---|---|---|---|---|---|---|---|---|
| unsloth/GLM-5.3-GGUF | quantized | GGUF | — | — | other | unsloth | 109 | 975.1k |
| nvidia/GLM-5.3-NVFP4 | quantized | NVFP4 | 391B | XXL | other | nvidia | 47 | 56.5k |
| amd/GLM-5.3-Quark-MXFP4-AttnFP8 | quantized | MXFP4 | 384B | XXL | other | amd | 4 | 27.8k |
| incoai/GLM-5.3-NVFP4 | quantized | NVFP4 | 391B | XXL | other | incoai | 26 | 20.6k |
| Justvugg/GLM-5.3-colibri-int4-g64 | quantized | int4 | — | — | other | Justvugg | 11 | 20.1k |
| incoai/GLM-5.3-DFlash2 | finetune | — | 2.46B | M | cc-by-nc-nd-4.0 | incoai | 25 | 15.7k |
| autotrust/GLM-5.3-GGUF-DGX-Spark | quantized | GGUF | — | — | other | autotrust | 19 | 9.9k |
| dealignai/GLM-5.3-UNCENSORED-NVFP4 | quantized | NVFP4 | 391B | XXL | mit | dealignai | 69 | 6.4k |
| orcarouter/GLM-5.3-MLX | quantized | MLX | 804B | XXL | other | orcarouter | 13 | 7.6k |
| dealignai/GLM-5.3-ABLITERATED-NVFP4 | quantized | NVFP4 | 391B | XXL | mit | dealignai | 49 | 3k |
| OneNexus/GLM-5.3-MXFP4 | quantized | MXFP4 | 390B | XXL | other | OneNexus | 4 | 5.2k |
| RedHatAI/GLM-5.3-MXFP4 | quantized | MXFP4 | 753B | XXL | other | RedHatAI | 3 | 4.9k |
| RedHatAI/GLM-5.3-NVFP4 | quantized | NVFP4 | 753B | XXL | other | RedHatAI | 4 | 4.4k |
| Tech2wild/GLM-5.3-Int4-Int8Mix | quantized | Int4 | 779B | XXL | other | Tech2wild | 16 | 3.2k |
| davidsyoung/GLM-5.3-EXL3-TR3-3.42bpw | quantized | EXL3 | 178B | XXL | other | davidsyoung | 11 | 3.1k |
| gpustack/GLM-5.3-W4A8 | quantized | — | 753B | XXL | other | gpustack | 4 | 2.8k |
| 0xSero/GLM-5.3-500B-EXL3-3.0bpw | quantized | EXL3 | 98.5B | XL | other | 0xSero | 29 | 1.5k |
| RedHatAI/GLM-5.3-speculator.dspark | finetune | — | 1.25B | S | other | RedHatAI | 12 | 1.9k |
| pipenetwork/GLM-5.3-MLX-mixed-4_8bit | quantized | MLX | 743B | XXL | other | pipenetwork | 2 | 2.1k |
| davidsyoung/GLM-5.3-EXL3-TR3-3.25bpw | quantized | EXL3 | 170B | XXL | other | davidsyoung | 7 | 1.8k |
| unsloth/GLM-5.3 | finetune | — | 753B | XXL | other | unsloth | 11 | 1.5k |
| cyankiwi/GLM-5.3-AWQ-INT4 | quantized | AWQ | 152B | XXL | other | cyankiwi | 2 | 1.8k |
| drowzeys/keys-GLM-5.3-EXL3-2.75BPW | quantized | EXL3 | 138B | XXL | other | drowzeys | 4 | 1.7k |
| inference-optimization/GLM-5.3-0.6B-A0.4B | finetune | — | 622M | S | mit | inference-optimization | 1 | 1.5k |
| DevQuasar/zai-org.GLM-5.3-GGUF | quantized | GGUF | — | — | — | DevQuasar | 2 | 1.5k |
| pipenetwork/GLM-5.3-MLX-4bit | quantized | MLX | 743B | XXL | other | pipenetwork | 2 | 1.5k |
| 0xSero/GLM-5.3-EXL3-3.0bpw | quantized | EXL3 | 146B | XXL | other | 0xSero | 9 | 1.1k |
| local-inference-lab/GLM-5.3-NVFP4 | quantized | NVFP4 | 391B | XXL | mit | local-inference-lab | 4 | 1.3k |
| Coder40-95/GLM-5.3-GGUF | quantized | GGUF | — | — | other | Coder40-95 | 0 | 1.5k |
| pipenetwork/GLM-5.3-MLX-mixed-3_6bit | quantized | MLX | 743B | XXL | other | pipenetwork | 1 | 1.4k |
| jarrelscy/GLM-5.3-Vision-NestQuant-2-4bit | quantized | -4bit | 29.0B | L | mit | jarrelscy | 0 | 1.5k |
| LibertAIDAI/GLM-5.3-NVFP4 | quantized | NVFP4 | 391B | XXL | other | LibertAIDAI | 1 | 1.3k |
| drowzeys/keys-GLM-5.3-EXL3-Abliterated | quantized | EXL3 | 165B | XXL | mit | drowzeys | 20 | 315 |
| davidsyoung/GLM-5.3-EXL3-TR3-3.0bpw | quantized | EXL3 | 158B | XXL | other | davidsyoung | 3 | 953 |
| pipenetwork/GLM-5.3-MLX-5bit | quantized | MLX | 743B | XXL | other | pipenetwork | 0 | 1.1k |
| pipenetwork/GLM-5.3-MLX-6bit | quantized | MLX | 743B | XXL | other | pipenetwork | 1 | 1k |
| Mia-AiLab/GLM-5.3-EXL3-2.75bpw-TensorFold | quantized | EXL3 | 146B | XXL | other | Mia-AiLab | 8 | 677 |
| JANGQ-AI/GLM-5.3-W4A16 | quantized | -W4A16 | 753B | XXL | other | JANGQ-AI | 2 | 940 |
| SixVolts/GLM-5.3-ewaste-edition-GGUF | quantized | GGUF | — | — | mit | SixVolts | 4 | 830 |
| vogel61/GLM-5.3-oQ4e | quantized | — | 743B | XXL | other | vogel61 | 1 | 970 |
| Date | Commit | SHA |
|---|---|---|
| 2026-10-14 | chat template: speed up tool-result validation and reordering (#31) | ba32f8c |
| 2026-10-12 | Add Agents' Last Exam result (28.5, ale-cli-pass-rate, from model card) (#30) | c5087a7 |
| 2026-09-04 | chat template: early break in tool result reordering check | aca966e |
| 2026-09-04 | Add Toolathlon-Verified evaluation result (#15) | 5bcee3f |
| 2026-09-04 | Point terminal-bench-3.0 eval result to unified harborframework/terminal-bench dataset (#16) | 5a1460b |
| 2026-08-31 | Move harness note into notes field for Terminal-Bench 2.1 and 3.0 (#13) | 187fb9f |
| 2026-08-31 | Add harness used for Terminal-Bench 2.1 and 3.0 eval results (#11) | eec49ed |
| 2026-08-29 | Add community evaluation results for HLE (#3) | e0b07fd |
| 2026-08-28 | Add community evaluation results for DEEP-SWE, TERMINAL-BENCH-2.1, TERMINAL-BENCH-3.0 (#2) | 935644c |
| 2026-08-28 | update | 3033303 |
| 2026-08-27 | Initial commit 0828 | 7cda819 |
Hosted endpoints observed for GLM-5.3, 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: $0.07/M.
| Provider | Model ref | Context | Quant | $/M in | $/M out | tok/s | TTFT | Tools | Status | Observed |
|---|---|---|---|---|---|---|---|---|---|---|
| Relace openrouter | z-ai/glm-5.3 | 1M | unknown | $0.07 | $4.29 | — | — | yes | live | 2026-10-24 |
| InferenceNet openrouter | z-ai/glm-5.3 | 1M | fp4 | $0.08 | $4.00 | — | — | yes | live | 2026-10-24 |
| Wafer openrouter | z-ai/glm-5.3 | 1M | fp4 | $0.09 | $1.61 | — | — | yes | live | 2026-10-24 |
| Reka openrouter | z-ai/glm-5.3 | 1M | unknown | $0.14 | $4.20 | — | — | yes | live | 2026-10-24 |
| Makora openrouter | z-ai/glm-5.3 | 1M | fp4 | $0.14 | $4.40 | — | — | yes | live | 2026-10-24 |
| AkashML openrouter | z-ai/glm-5.3 | 1M | fp8 | $0.19 | $4.40 | — | — | yes | live | 2026-10-24 |
| DekaLLM openrouter | z-ai/glm-5.3 | 1M | unknown | $0.25 | $2.95 | — | — | yes | live | 2026-10-24 |
| DeepInfra hf, deepinfra, openrouter | zai-org/GLM-5.3 | 1M | fp4 | $0.45 | $2.00 | 73.7 | 3688 ms | yes | live | 2026-10-25 |
| Morph openrouter | z-ai/glm-5.3 | 1M | fp8 | $0.45 | $4.40 | — | — | yes | live | 2026-10-24 |
| Sail Research openrouter | z-ai/glm-5.3 | 1M | fp8 | $0.50 | $2.00 | — | — | yes | live | 2026-10-24 |
| Baidu openrouter | z-ai/glm-5.3 | 1M | fp8 | $0.60 | $1.88 | — | — | yes | live | 2026-10-24 |
| Tenstorrent openrouter | z-ai/glm-5.3 | 1020K | fp8 | $0.60 | $2.50 | — | — | yes | live | 2026-10-24 |
| Inceptron openrouter | z-ai/glm-5.3 | 1M | fp4 | $0.60 | $3.22 | — | — | yes | live | 2026-10-24 |
| Novita hf, openrouter | zai-org/glm-5.3 | 1M | fp8 | $0.63 | $1.98 | 70.7 | 1014 ms | yes | live | 2026-10-25 |
| SiliconFlow openrouter | z-ai/glm-5.3 | 1M | fp8 | $0.70 | $2.20 | — | — | yes | live | 2026-10-24 |
| GMICloud openrouter | z-ai/glm-5.3 | 1M | fp8 | $0.70 | $2.20 | — | — | yes | live | 2026-10-24 |
| Phala openrouter | z-ai/glm-5.3 | 1M | unknown | $0.84 | $2.64 | — | — | yes | live | 2026-10-24 |
| Decart openrouter | z-ai/glm-5.3 | 1M | fp4 | $0.84 | $2.71 | — | — | yes | live | 2026-10-24 |
| DigitalOcean openrouter | z-ai/glm-5.3 | 1M | unknown | $0.91 | $2.86 | — | — | yes | live | 2026-10-24 |
| Fireworks AI hf, openrouter | accounts/fireworks/models/glm-5p3 | 1M | unknown | $0.98 | $3.08 | 46.1 | 557 ms | yes | live | 2026-10-25 |
| PrimeIntellect openrouter | z-ai/glm-5.3 | 1M | unknown | $0.98 | $3.08 | — | — | yes | live | 2026-10-24 |
| Alibaba openrouter | z-ai/glm-5.3 | 977K | unknown | $1.19 | $3.74 | — | — | yes | live | 2026-10-24 |
| CoreWeave openrouter | z-ai/glm-5.3 | 1M | fp8 | $1.20 | $4.00 | — | — | yes | live | 2026-10-24 |
| Friendli openrouter | z-ai/glm-5.3 | 1M | unknown | $1.26 | $3.96 | — | — | yes | live | 2026-10-24 |
| Together AI hf, openrouter | zai-org/GLM-5.3 | 1024K | unknown | $1.40 | $4.40 | 88.8 | 329 ms | yes | live | 2026-10-25 |
| Z.ai hf, openrouter | glm-5.3 | 1M | fp8 | $1.40 | $4.40 | 50.9 | 2546 ms | yes | live | 2026-10-25 |
| Baseten hf, openrouter | zai-org/GLM-5.3 | 1M | fp4 | $1.40 | $4.40 | 94.8 | 470 ms | yes | live | 2026-10-25 |
| Parasail parasail, openrouter | parasail-glm-53 | 1M | fp8 | $1.40 | $4.40 | — | — | yes | serverless | 2026-10-24 |
| Mistral openrouter | z-ai/glm-5.3 | 1M | nvfp4 | $1.40 | $4.40 | — | — | yes | live | 2026-10-24 |
| Nebius openrouter | z-ai/glm-5.3 | 1000K | fp4 | $1.40 | $4.40 | — | — | yes | live | 2026-10-24 |
| Crusoe openrouter | z-ai/glm-5.3 | 1M | fp4 | $1.40 | $4.40 | — | — | yes | live | 2026-10-24 |
| Venice openrouter | z-ai/glm-5.3 | 977K | unknown | $1.40 | $4.40 | — | — | yes | live | 2026-10-24 |
| Modal openrouter | z-ai/glm-5.3 | 1M | unknown | $1.40 | $4.40 | — | — | yes | live | 2026-10-24 |
| Cloudflare openrouter | z-ai/glm-5.3 | 1M | unknown | $1.40 | $4.40 | — | — | yes | live | 2026-10-24 |
| AtlasCloud openrouter | z-ai/glm-5.3 | 1M | fp8 | $1.40 | $4.40 | — | — | yes | live | 2026-10-24 |
| Amazon Bedrock openrouter | z-ai/glm-5.3 | 1M | unknown | $1.68 | $5.28 | — | — | yes | live | 2026-10-24 |
| ModelRun openrouter | z-ai/glm-5.3 | 1M | fp4 | $2.10 | $6.60 | — | — | yes | live | 2026-10-24 |
| Featherless hf | zai-org/GLM-5.3 | — | — | — | — | — | — | — | live | 2026-10-25 |
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.
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).
Hub snapshot: 2026-10-25 · HF API: 2026-10-24 21:31 UTC · Providers: 2026-10-24 21:44 UTC · Articles: 2026-10-24 21:45 UTC · Summary: 2026-10-07 14:59 UTC · Last verified 2026-10-25
Email alerts are not open yet; the feeds above are the same data. See feeds for the item schema.