GLM-5.2

Fiche modèle · Zhipu AI · glm-5.2

GLM-5.2 is Z.AI's flagship foundation model built for long-horizon tasks, delivering a solid 1M-token context that stably sustains long, messy coding-agent trajectories. It improves substantially over GLM-5.1, becoming the strongest open-source model on standard coding benchmarks (81.0 on Terminal-Bench 2.1 and 62.1 on SWE-bench Pro) and the highest-ranked open-source model across long-horizon coding benchmarks (FrontierSWE, PostTrainBench, SWE-Marathon). It introduces flexible thinking effort levels (High and Max) to balance capability against latency and compute. Architecturally, GLM-5.2 proposes IndexShare, which reuses one lightweight indexer across every four sparse-attention (DSA) layers to cut per-token FLOPs by 2.9x at 1M context, and an improved MTP layer for speculative decoding that raises acceptance length by up to 20%. Released under a pure MIT open-source license with weights available on HuggingFace and ModelScope, it supports transformers, vLLM, SGLang, xLLM, and ktransformers, with 1M input context, 128K max output, thinking mode, function calling, structured output, context caching, and MCP integration.

Identité

OrganisationZhipu AI
Pays—
Famille—
Typellm
Date de sortie2026-06-16
Licencemit
Licence commercialeoui
Licence open sourceoui
Poids ouvertsoui
Auto-hébergement (DGX Spark)non — 753 Md > 107 Md
Modalitéstext
Fenêtre de contexte1 048 576
KV cache / token (Q8)1.92 Mo/token
Paramètres totaux753 B
Paramètres actifs (session)40 B
Source des paramètres actifsweb : https://github.com/zai-org/GLM-5
Curation des paramètres actifs2026-11-04
Prix par session1.23 $ (1 000 000 tokens entrée + 200 000 sortie, prix le moins cher par fournisseur)
Tokens / tâche (Data Players)1.2 M
Tokens / tâche (Terminal-Bench)83.3 M (est.)
Ratio Terminal-Bench / Data Players69.4×
Coût / tâche (Terminal-Bench)10.33 $ (est.)
Tokens d'entraînement—
Coupure de connaissance—
Créé le2026-06-16T19:03:52.341474Z
Mis à jour le2026-11-04T06:05:14.076186Z
SourceLLM Stats
URLhttps://llm-stats.com/models/glm-5.2

IA Act — preuves fournisseur

ICA 2/100 · open source · seuil systémique inconnu

CritèreNote / 5Poids
Documentation technique2,55 %
Résumé des données d'entraînement0,035 %
Conformité droit d'auteur0,035 %
Code of Practice GPAI0,020 %
Risque systémique0,05 %

Curation : 2026-10-23.

IA Act

Prix par fournisseur

Statut du modèle (offres + auto-hébergement) : souverain — via Arkane (cortecs), GreenPT (eurouter), Inceptron (cortecs, eurouter), Lyceum Technology GmbH (eurouter), Mistral AI (cortecs), Nebul (eurouter), Regolo (eurouter), Scaleway (cortecs, eurouter, scaleway), Tensorix (cortecs, eurouter).

Toutes sources confondues : catalogue llm-stats, routeurs EUrouter/Cortecs et catalogues directs (OVHcloud, Scaleway, Infomaniak). Les prix sont convertis en USD pour la comparaison ; le prix publié et le taux BCE utilisé sont dans l'infobulle. Les prix inconnus restent indiqués « — ».

FournisseurSourceSouverainetéEntrée (/M tok)Sortie (/M tok)QuantificationContexteLatence (s)Débit (tok/s)Statut
Arkanecortecssouverain1.31 $4.23 $—200 000———
GreenPTeuroutersouverain———1 048 576———
Inceptroncortecs, euroutersouverain0.91 $3.26 $fp41 000 000 / 1 048 576———
Lyceum Technology GmbHeuroutersouverain———1 048 576———
Mistral AIcortecssouverain1.47 $4.62 $—1 000 000———
Nebuleuroutersouverain———1 048 576———
Regoloeuroutersouverain———1 048 576———
Scalewaycortecs, eurouter, scalewaysouverain2.02 $6.17 $—256 000 / 1 048 576———
Tensorixcortecs, euroutersouverain1.49 $4.48 $fp81 048 576———
DeepInfrallm-statsnon0.75 $2.40 $————active
Fireworks AIllm-statsnon1.40 $4.40 $————active
FriendliAIllm-statsnon1.40 $4.40 $————active
Novitallm-statsnon1.40 $4.40 $————active
Together AIllm-statsnon1.40 $4.40 $————active
ZAIllm-statsnon1.40 $4.40 $————active

Souveraineté qualifiée par offre (modèle × fournisseur) ; survolez le badge pour les critères.

Souveraineté

Scores par catégorie (top_scores)

CatégorieScore
math0.70
physics0.90
reasoning0.60
general0.60
agents0.50
biology0.90
chemistry0.90
code0.50
systems0.30
tool_calling0.70
vision0.50

Scores par benchmark

BenchmarkCatégorieScoreMaxNormalisé (API)VérifiéAuto-déclaréDaté du
aime-2026math0.991.000.99nonoui2026-11-04T06:05:14.089180Z
critptreasoning0.171.000.17nonoui2026-11-04T06:05:14.086223Z
deepswecode0.461.000.46nonoui2026-11-04T06:05:14.110685Z
deepswe-1.1code0.441.000.44nonnon2026-11-04T06:05:14.142081Z
frontiercode-1.1reasoning0.241.000.24nonnon2026-11-04T06:05:14.145384Z
frontierswecode0.741.000.74nonnon2026-11-04T06:05:14.125922Z
gpqareasoning0.911.000.91nonoui2026-11-04T06:05:14.101435Z
hmmt-2025math0.941.000.94nonoui2026-11-04T06:05:14.092075Z
hmmt-feb-26math0.931.000.93nonoui2026-11-04T06:05:14.094949Z
humanity's-last-examreasoning0.551.000.55nonoui2026-11-04T06:05:14.082836Z
imo-answerbenchmath0.911.000.91nonoui2026-11-04T06:05:14.098156Z
mcp-atlasreasoning0.771.000.77nonoui2026-11-04T06:05:14.135709Z
nl2repocode0.491.000.49nonoui2026-11-04T06:05:14.107224Z
posttrainbenchagents0.341.000.34nonoui2026-11-04T06:05:14.129341Z
program-benchagents0.641.000.64nonoui2026-11-04T06:05:14.113638Z
swe-bench-proreasoning0.621.000.62nonoui2026-11-04T06:05:14.104251Z
swe-marathoncode0.131.000.13nonoui2026-11-04T06:05:14.132481Z
terminal-bench-2.1reasoning0.831.000.83nonoui2026-11-04T06:05:14.119819Z
toolathlonagents0.481.000.48nonoui2026-11-04T06:05:14.138730Z

Classements

CatégorieRangScoreNote conservatriceBenchmarks évaluésMéthodeDaté duProvenance
code470.5033.7611trueskill2026-11-04T08:08:39.742311Zrankings
math110.7041.476trueskill2026-11-04T08:08:42.442591Zrankings
reasoning480.6044.4012trueskill2026-11-04T08:08:41.093494Zrankings
agents52—28.7812trueskill2026-10-20model_history

Lecture : Note conservatrice = borne basse TrueSkill (niveau estimé − marge d’incertitude) ; Benchmarks évalués = nombre de benchmarks de la catégorie pris en compte. À performances égales, un modèle évalué sur moins de benchmarks obtient une note plus basse : la note ne se déduit pas d’un seul score. Provenance : rankings = top 50 de /rankings ; model_history = historique du modèle (/models/{id}/history), utilisé pour les rangs au-delà du top 50.

Données

Dates de curation (récupération depuis l'API) : models 2026-11-04 · benchmarks 2026-11-04 · scores 2026-11-04 · updates 2026-11-04 · account 2026-11-04 · model_history 2026-11-04 · rankings 2026-11-04