GLM-5 / GLM-5.1

Overview

GLM-5 is the flagship open-weights large language model of Zhipu AI (operating internationally as Z.ai, Hugging Face org zai-org), launched "targeting complex systems engineering and long-horizon agentic tasks" (GLM-5 model card). The GLM-5 weight repository was created on Hugging Face on 2026-02-11 (Hugging Face model API for zai-org/GLM-5), and secondary coverage dates the public release to February 11, 2026 (NYU Shanghai RITS). Compared with GLM-4.5, GLM-5 scales from 355B parameters (32B active) to 744B parameters (40B active) and raises pre-training data from 23T to 28.5T tokens (GLM-5 model card; Z.ai GLM-5 blog). Weights are distributed under the MIT license (GLM-5 README; Hugging Face model API), on Hugging Face and ModelScope, with API access on api.z.ai and BigModel.cn (Z.ai GLM-5 blog).

GLM-5.1 is the follow-up point release on the same architecture, described by Z.ai as "our next-generation flagship model for agentic engineering, with significantly stronger coding capabilities than its predecessor," reaching state-of-the-art performance on SWE-Bench Pro and leading GLM-5 by a wide margin on NL2Repo and Terminal-Bench 2.0 (GLM-5.1 model card). Its distinguishing claim is long-horizon durability: where earlier models "exhaust their repertoire early," GLM-5.1 "sustains optimization over hundreds of rounds and thousands of tool calls" (GLM-5.1 model card). The GLM-5.1 repository was created 2026-04-03 (Hugging Face model API for zai-org/GLM-5.1) and third-party documentation gives the public release date as April 7, 2026 (Clore.ai GLM-5.1 guide); it is also MIT-licensed (GLM-5.1 README; Z.ai GLM-5.1 blog). Both models are served through the Z.ai API with a 200K context window and 128K maximum output (Z.ai docs: GLM-5; Z.ai docs: GLM-5.1).

Notable techniques: GLM-5 integrates DeepSeek Sparse Attention (DSA), "largely reducing deployment cost while preserving long-context capacity" (GLM-5 model card), and the arXiv technical report GLM-5: from Vibe Coding to Agentic Engineering (arXiv:2602.15763, submitted 17 Feb 2026) states that GLM-5 "adopts DSA" to cut training and inference cost while maintaining long-context fidelity, adds an asynchronous RL infrastructure that decouples generation from training, and proposes novel asynchronous agent RL algorithms (arXiv:2602.15763). That infrastructure is named slime, "a novel asynchronous RL infrastructure that substantially improves training throughput and efficiency, enabling more fine-grained post-training iterations" (GLM-5 model card). GLM-5.1 cites the same technical report as GLM-5, consistent with an architecture-sharing point release (GLM-5.1 model card). Beyond NVIDIA, Z.ai reports support on Huawei Ascend, Moore Threads, Cambricon, Kunlun Chip, MetaX, Enflame and Hygon accelerators (Z.ai GLM-5 blog).

Architecture

Values below come from the published config.json of each repository; GLM-5 and GLM-5.1 configs are byte-for-byte identical apart from the transformers_version field, confirming a shared architecture.

• Total parameters (vendor figure): 744B (GLM-5 model card; Z.ai docs: GLM-5)
• Total parameters (checkpoint tensor count): 753,864,139,008 (~754B) for both GLM-5 and GLM-5.1 (GLM-5 HF API; GLM-5.1 HF API)
• Active parameters per token: 40B (GLM-5 model card; Z.ai docs: GLM-5)
• Model class / architecture string: GlmMoeDsaForCausalLM, model_type: glm_moe_dsa (GLM-5 config.json; GLM-5.1 config.json)
• Layers: 78 hidden layers, with the first 3 layers dense (first_k_dense_replace: 3) and MoE in every subsequent layer (moe_layer_freq: 1) (GLM-5 config.json)
• Hidden size / dense FFN: hidden_size 6144, intermediate_size 12288 (GLM-5 config.json)
• MoE experts: 256 routed experts + 1 shared expert, 8 experts activated per token, expert FFN width 2048, sigmoid scoring with noaux_tc top-k routing and routed scaling factor 2.5 (GLM-5 config.json; GLM-5.1 config.json)
• Attention: DeepSeek Sparse Attention (DSA) (GLM-5 model card; arXiv:2602.15763); MLA-style low-rank projections with q_lora_rank: 2048, kv_lora_rank: 512, 64 attention heads, qk_head_dim: 256 (192 nope + 64 rope), v_head_dim: 256 (GLM-5 config.json)
• DSA indexer: 32 indexer heads, index_head_dim: 128, index_topk: 2048 selected tokens per query, interleaved indexer RoPE (GLM-5 config.json); in DSA a lightweight "lightning indexer" picks the top-k relevant tokens, cutting core attention from O(L²) to O(Lk) (IndexCache paper, arXiv:2603.12201)
• Multi-token prediction (MTP): num_nextn_predict_layers: 1, used for speculative decoding (vLLM launch uses --speculative-config.method mtp, SGLang uses EAGLE) (GLM-5 config.json; GLM-5 model card)
• Context window: max_position_embeddings: 202752 in the checkpoint (GLM-5 config.json); Z.ai API lists a 200K context window and 128K max output for both GLM-5 and GLM-5.1 (Z.ai docs: GLM-5; Z.ai docs: GLM-5.1)
• RoPE: default RoPE with theta 1,000,000, interleaved (GLM-5 config.json)
• Precision: BF16 checkpoint (dtype: bfloat16), plus an FP8 sibling repository zai-org/GLM-5.1-FP8 listing F8_E4M3 tensors at 754B parameters for cheaper serving (GLM-5 config.json; zai-org/GLM-5.1-FP8); GLM-5.1 tensors are listed as BF16/F32 (GLM-5.1 model card)
• Vocabulary: 154,880 tokens, untied embeddings (GLM-5 config.json)
• Pre-training data: 28.5T tokens (up from 23T for GLM-4.5) (GLM-5 model card)
• License: MIT (GLM-5 README; GLM-5.1 README)
• Serving stacks: vLLM v0.19.0+, SGLang v0.5.10+, KTransformers v0.5.3+, Transformers, xLLM v0.8.0+ (GLM-5 model card; GLM-5.1 model card)

Official vendor eval scores

Vendor-reported only (Hugging Face model cards and Z.ai blog). GLM-5 column values are from the GLM-5 card/blog; GLM-5.1 values from the GLM-5.1 card/blog.

Notes on evaluation harnesses (vendor-stated): SWE-bench runs use OpenHands with a 200K context window; Terminal-Bench 2.0 Claude Code scores are averaged over 5 runs; CyberGym is single-run Pass@1 over 1,507 tasks; Vending Bench 2 runs were conducted independently by Andon Labs (GLM-5 model card). GLM-5's τ²-Bench airline domain uses the fixes from the Claude Opus 4.5 system card (GLM-5 model card). Where the GLM-5 and GLM-5.1 cards disagree on a GLM-5 number (CyberGym, MCP-Atlas, AIME, Tool-Decathlon), both figures are shown above.

Benchmark explanations

• SWE-bench Verified — human-validated GitHub issues that the model must resolve by patching a real repository; measures end-to-end software bug-fixing (GLM-5 model card).
• SWE-bench Multilingual — the same issue-resolution task across repositories in multiple programming languages, testing generality beyond Python (GLM-5 model card).
• SWE-Bench Pro — a harder, contamination-resistant issue-resolution suite (hosted by Scale AI) run here with the OpenHands agent under a 400K/200K context (GLM-5.1 model card).
• NL2Repo — repository generation: build a working codebase from a natural-language specification (GLM-5.1 model card).
• Terminal-Bench 2.0 — real-world terminal tasks executed in a sandboxed shell under CPU/RAM limits, scored by task completion (GLM-5.1 model card; GLM-5 model card).
• τ²-Bench / τ³-Bench — multi-turn tool-use and customer-service style agent dialogues in Retail, Telecom and Airline domains, testing policy compliance with a simulated user (GLM-5 model card).
• BrowseComp / BrowseComp-Zh — hard web-browsing research questions requiring many search-and-read hops; the Zh variant is the Chinese-language version, with and without context-management strategies (GLM-5 model card).
• AIME 2026 — American Invitational Mathematics Examination problems, short-answer competition math (GLM-5.1 model card).
• HMMT (Nov. 2025 / Feb. 2026) — Harvard-MIT Mathematics Tournament problem sets, competition-level math reasoning (GLM-5.1 model card).
• IMOAnswerBench — International-Mathematical-Olympiad-style problems graded on final answers (GLM-5.1 model card).
• GPQA-Diamond — the hardest subset of graduate-level, Google-proof physics/chemistry/biology multiple-choice questions (GLM-5.1 model card).
• Humanity's Last Exam (HLE) — a very broad expert-written exam of frontier knowledge and reasoning; reported text-only by default and separately with tool access (GLM-5 model card).
• CyberGym — starts from white-box source code and tests whether the model can identify and validate vulnerabilities by triggering faults (Z.ai GLM-5.3 blog); evaluated as single-run Pass@1 over 1,507 tasks (GLM-5 model card).
• MCP-Atlas — 500-task public subset of tool-calling tasks over Model Context Protocol servers, LLM-judged (GLM-5 model card).
• Tool-Decathlon — a ten-domain battery of tool-use/agent tasks run through an official evaluation service (GLM-5.2 model card).
• Vending Bench 2 — a simulated vending-machine business run over a one-year horizon, scored by final account balance; measures long-term planning and resource management (Z.ai GLM-5 blog).
• VectorDBBench / KernelBench (GLM-5.1 case studies) — building a high-QPS approximate-nearest-neighbour database in Rust (GLM-5.1: 21.5k QPS over 600+ iterations and 6,000+ tool calls) and generating faster GPU kernels from PyTorch references (GLM-5.1: 3.6× geometric-mean speedup over 50 problems) (Z.ai GLM-5.1 blog).