GLM-5.2

GLM-5.2 is an open-weight large language model developed by the Chinese company Zhipu AI, which sells its products internationally under the brand Z.ai. Released in June 2026 under the permissive MIT License, it is a Mixture-of-Experts (MoE) model that Z.ai designates 744B-A40B, meaning about 744 billion total parameters and about 40 billion active per token, a 1 million token context window, and a focus on agentic coding and long-horizon task execution. At launch it became the highest-scoring open-weight model on the Artificial Analysis Intelligence Index, and on Z.ai's own reported evaluations it out-scored OpenAI's GPT-5.5 on several long-horizon coding benchmarks at a fraction of the price. ^[1][2][3] The US government's Center for AI Standards and Innovation later concluded, after running its own evaluations, that GLM-5.2 "was probably the most capable open-weight AI model when it was released," while placing its overall capability roughly level with OpenAI's GPT-5.2 from December 2025. ^[9]

Scroll sideways for more →

What is GLM-5.2?

GLM-5.2 is the third release in Zhipu AI's fifth-generation GLM (General Language Model) line, following GLM-5 and GLM-5.1. It is built for "long-horizon" work: multi-step reasoning, tool use, and software-engineering tasks that unfold over many turns and large codebases. Z.ai describes it as a substantial leap in long-horizon capability over GLM-5.1, delivered for the first time on a full 1 million token context. ^[2] The weights are published on Hugging Face and ModelScope under the repository zai-org/GLM-5.2, so anyone can download, run, fine-tune, and deploy the model commercially. ^[3][13] Z.ai's developer documentation lists the model as text in, text out, with a 1M token context and a 128K token maximum output, and exposes a reasoning-effort parameter that includes a max setting. ^[5]

The rollout ran in stages over four days. Subscribers to Z.ai's GLM Coding Plan got access on June 13, 2026; the metered API and the open weights both landed on June 16; and Z.ai's technical write-up went up the following day. ^[6][9][2] There is no technical report for GLM-5.2 as a model. The architectural background is documented in the GLM-5 team's paper, "GLM-5: from Vibe Coding to Agentic Engineering" (arXiv 2602.15763), and the main 5.2-specific change, the shared sparse-attention indexer, has a paper of its own: "IndexCache: Accelerating Sparse Attention via Cross-Layer Index Reuse" (arXiv 2603.12201), which the model card links from the word IndexShare. The remaining 5.2 changes are described only in Z.ai's blog post. ^{[19][3][25][2]}

Who developed GLM-5.2?

GLM-5.2 was built by Zhipu AI, a Beijing-based AI company founded in 2019 as a spin-off from Tsinghua University's Knowledge Engineering Group. The company markets its models and API internationally under the Z.ai brand and trades in Hong Kong as Knowledge Atlas Technology (HKEX: 2513) after completing its IPO on January 8, 2026, which made it the first major Chinese large-model developer to go public. CAISI recorded Z.ai's market capitalization as roughly $7 billion at IPO and around $65 billion by July 17, 2026, on 2025 revenue of $105 million. ^[9][8] GLM-5.2 continues Zhipu's strategy of shipping frontier-class models with open weights, positioning the firm as a direct competitor to both Western labs and other Chinese open-weight developers.

How does GLM-5.2 relate to GLM-5 and GLM-5.1?

The series began with GLM-5, a 744 billion parameter open-weight flagship released on February 11, 2026. GLM-5.1 followed on April 7, 2026 as an incremental update. GLM-5.2 is the coding and agent focused refinement of that lineage. It keeps the MoE design but raises the context window from 200,000 tokens to 1 million and adds architectural changes aimed at cheaper long-context inference. Z.ai reports large coding gains over GLM-5.1, for example Terminal-Bench 2.1 rising to 81.0 from 63.5. ^[2][1]

Some of the biggest reported jumps are on benchmarks where GLM-5.1 barely registered. On FrontierSWE, which scores agents on open-ended projects that run for hours, Z.ai reports 74.4 for GLM-5.2 against 30.5 for GLM-5.1; on SWE-Marathon it reports 13.0 against 1.0. ^[2] Large deltas of that kind usually mean the older model was failing the task format outright rather than solving a slightly harder version of the same problem, so they are better read as a capability threshold being crossed than as a linear improvement.

What is GLM-5.2's architecture?

According to its Hugging Face config.json, GLM-5.2 uses the GlmMoeDsaForCausalLM architecture (model type glm_moe_dsa). It has 78 transformer layers, a hidden size of 6,144, and 64 attention heads. The MoE feed-forward network contains 256 routed experts plus 1 shared expert, with 8 routed experts activated per token, giving roughly 40 billion active parameters. The vocabulary is 154,880 tokens and the weights ship in bfloat16. ^[4][2]

Two different totals circulate for this model, and both are real. Z.ai's own GitHub model index labels every GLM-5 series checkpoint "744B-A40B," and Artificial Analysis's launch write-up gives "744B total parameters, 40B active parameters"; adding up the layer shapes in config.json reproduces roughly that number. Hugging Face instead displays 753B, because the safetensors metadata counts the entire published checkpoint, 753,329,921,024 parameters, and that includes the draft block the config's num_nextn_predict_layers setting adds for speculative decoding. The 9.6 billion difference is about the size of one more of the model's MoE blocks, which is what that block is. Artificial Analysis carries both figures on different pages, so a number seen in isolation says nothing about which basis it uses. ^{[13][1][4][28]}

The config also shows the first 3 layers are dense and the remaining 75 are sparse, that attention is DeepSeek-style multi-head latent attention rather than grouped-query attention, with a 512-dimension compressed key-value latent, a 2,048-dimension query latent, and 64 nominal key-value heads matching the 64 query heads, and that RoPE theta is set to 8,000,000 to support the 1,048,576 token position range. A sparse-attention indexer with 32 heads and a top-k of 2,048 governs which past tokens each layer attends to, and a single next-token prediction layer supports speculative decoding. ^[4]

IndexShare and multi-token prediction

Two efficiency features distinguish the model. "IndexShare" reuses a single lightweight sparse-attention indexer across every four transformer layers, which Z.ai says cuts per-token FLOPs by about 2.9 times at the full 1 million token context. An upgraded multi-token prediction layer improves speculative decoding: in an ablation Z.ai ran on GLM-5.1's backbone and training data, the combination of IndexShare, KV sharing, rejection sampling, and an end-to-end TV loss raised the acceptance length from 4.56 to 5.47, about 20 percent. Together with the DeepSeek-style sparse attention implied by the architecture name, these changes target the cost of very long context inference. ^[2]

Z.ai also describes the post-training pipeline in outline: it used the slime framework to run what it calls parallel OPD training, merging more than ten expert models in a process that took roughly two days, and it added anti-hack mechanisms combining rule-based filtering with LLM judging to suppress reward hacking. No training data disclosure accompanies these claims. ^[2]

Is GLM-5.2 open source?

Yes, in the open-weight sense, and with unusually few strings attached. The LICENSE file in the Hugging Face weights repository is the plain, unmodified MIT License: it requires only that the copyright and permission notice travel with copies, and imposes no field-of-use restriction, no regional limit, and no user-count or revenue threshold. ^[14] That distinction matters in mid-2026, because several competing Chinese open-weight releases ship modified licenses. Kimi K2.7-Code, which reached Hugging Face in the same week, uses a "Modified MIT License" that adds an attribution condition above a usage threshold. ^[21] Z.ai's GitHub repository for the GLM-5 family is separately licensed under Apache 2.0, which covers the inference and evaluation code rather than the weights. ^[15]

You can self-host by downloading the weights from Hugging Face or ModelScope. Z.ai ships two official checkpoints, the BF16 release and an FP8 version, and lists support for vLLM (0.23.0 and later), SGLang (0.5.13.post1 and later), Transformers, KTransformers, Unsloth, and Ascend NPU platforms. ^[13][16] As with most open-weight models, the weights and license are public, but the full training data and pipeline are not. ^[3]

How does GLM-5.2 perform on benchmarks?

The table below reproduces Z.ai's own published evaluation results. These are vendor-reported figures, produced by the model's developer with its own harnesses and prompts, and several of the entries for rival models are copied from those vendors' own reports rather than re-run by Z.ai. Treat them as claims, not measurements.

Scroll sideways for more →

Source: Z.ai model card. Because GLM-5.2 has no image input, its reasoning results are computed on text-only subsets, while the Opus 4.8 and GPT-5.5 Humanity's Last Exam entries are full-set numbers, so those two rows are not strictly like-for-like. ^[3][2]

Several of the benchmarks in the coding block are worth reading carefully. FrontierSWE, PostTrainBench, SWE-Marathon, NL2Repo, and DeepSWE are not part of the standard public evaluation suite that most labs report, and the "best reported harness" variant of Terminal-Bench lets each model use a different agent scaffold. GLM-5.2 does not appear on the official Terminal-Bench 2.1 leaderboard at all, whose seventeen entries as of 2026-07-27 were each run and verified by a Terminal-Bench team member rather than self-reported; the only Z.ai entry there is GLM-5.1 under Claude Code at max effort, at 58.7 percent. A benchmark suite chosen and run by the vendor is weaker evidence than a public benchmark run by the vendor, which is in turn weaker than a third-party run. ^[3][26]

How good is GLM-5.2 at coding?

Coding and agentic tool use are GLM-5.2's headline strengths. On Z.ai's numbers it scores 62.1 on SWE-bench Pro and 74.4 on FrontierSWE, beating GPT-5.5 (58.6 and 72.6) on both, and it reaches 81.0 on Terminal-Bench 2.1 with the standard Terminus-2 harness, roughly 18 points above GLM-5.1. VentureBeat's coverage led on the same comparison, framing the model as matching or beating GPT-5.5 on long-horizon coding for about one-sixth the price. ^[2][6] Z.ai's summary of its own position is that GLM-5.2 is "the strongest open-source model" on standard coding benchmarks, with "capability roughly positioned between Claude Opus 4.7 and Claude Opus 4.8 under similar token consumption." ^[9]

Independent evidence is more mixed. CAISI's re-run of SWE-Bench Verified put GLM-5.2 at 75.3 percent against 87.1 for Claude Opus 4.8 and 81.0 for GPT-5.5, and on PortBench, a non-public benchmark CAISI built itself so that no model could have trained on it, GLM-5.2 managed 41.7 percent against 78.0 for GPT-5.5 and 60.7 for Opus 4.8. CAISI's blunt reading was that it had "measured GLM-5.2's capabilities to be significantly lower than Opus 4.6's on an uncontaminated CAISI-created software-engineering benchmark," Claude Opus 4.6 being a model released four months earlier. CAISI publishes no PortBench score for Opus 4.6 itself, so that comparison rests on its aggregate capability model rather than on a printed number. ^[9] The gap between that result and Z.ai's own coding numbers is one of the most useful things known about this model.

What did independent evaluators measure?

Two independent bodies published results.

Artificial Analysis, in a write-up published on 2026-06-16, placed GLM-5.2 at 51 on Intelligence Index v4.1, the top score for any open-weight model at release, ahead of DeepSeek V4 Pro and MiniMax M3 (both 44) and Kimi K2.6 (43). It recorded the largest gains in scientific reasoning: CritPt up 16 points to 21 percent, Humanity's Last Exam up 12 points to 40 percent, and SciCode up 7 points to 50 percent. On its GDPval-AA v2 agentic evaluation GLM-5.2 scored 1,524, which Artificial Analysis called "effectively level with GPT-5.5 (xhigh, 1514)." Artificial Analysis also flagged the model's verbosity: about 43,000 output tokens per Index task, 37,000 of them reasoning, which works out to roughly $0.46 per task against $0.25 for GLM-5.1 and $0.05 for DeepSeek V4 Pro. ^[1]

The Center for AI Standards and Innovation (CAISI), part of NIST, published a 21-page assessment, dated July 8, 2026 and released on July 17, based on its own evaluation runs. Its top-line conclusion was that GLM-5.2 was probably the most capable open-weight model at release but still lagged the US frontier, with overall capability similar to GPT-5.2 from December 2025 and cyber capability similar to Claude Opus 4.6 from February 2026. ^[9][10]

Scroll sideways for more →

Source: CAISI, "Assessment of Z.ai's GLM-5.2," Table A1.1. DeepSeek's 32 on CTF-Archive Diamond is imputed from a subset of samples via item response theory rather than measured directly, and the full table also carries an Anthropic Mythos Preview column that leads most rows. CAISI notes two caveats that cut in opposite directions: its SWE-bench Verified scores run lower than other evaluators' because of system-prompt, scaffolding, and token-budget differences, and its mathematics runs used a custom output-format prompt taken from Z.ai's model card, without which scores fell 5 to 10 percent. ^[9] CAISI's GPQA Diamond figure of 91.4 sits within a rounding error of the 91.2 Z.ai reported, which is a genuine corroboration of at least one vendor number.

What are GLM-5.2's safety and security properties?

CAISI's assessment is the only published third-party security evaluation. It found a mixed picture. GLM-5.2 refused most overtly malicious cyber requests, complying with about 2 percent of them without a jailbreak and 8 percent with the best public jailbreak, comparable to Claude Opus 4.8 and better than DeepSeek V4 Pro, which complied with 100 percent under jailbreak. But on ten ExploitBench tasks involving agentic exploit development, GLM-5.2 never refused, running to the 300-turn limit without blocking, where Opus 4.8 blocked after 36 messages on average and Claude Opus 4.7 after 12. Two of the reference models behaved as GLM-5.2 did, however: GPT-5.5 and Claude Opus 4.6 also ran the full 300 turns without ever blocking. ^[9]

On sensitive biological queries GLM-5.2 was fully compliant 14 percent of the time without a jailbreak and 27 percent with one, lower than the two reference PRC models CAISI tested but far above the tested US models, which registered 0 percent in both conditions. ^[9]

Agent security was GLM-5.2's strongest showing. Against the best of 31 publicly available agent-hijacking attacks, its attack success rate was 0.0 percent, below Claude Opus 4.7 (0.1 percent), GPT-5.5 (0.5 percent), Kimi K2.6 (1.9 percent), and DeepSeek V4 Pro (15.1 percent). Under automated red-teaming that generated novel attacks, however, the success rate against GLM-5.2 climbed to roughly 80 percent after about 1,500 iterations, while GPT-5.5 and Opus 4.7 stayed near zero across the same budget. CAISI adds the structural caveat that applies to every open-weight release: safeguards measured through prompting are a lower bound, because anyone self-hosting the weights can strip refusal behavior outright. ^[9]

How does GLM-5.2 compare to other Chinese open-weight models?

Mid-2026 has been crowded. Among models whose weights have actually been published, GLM-5.2 still led the Artificial Analysis Intelligence Index v4.1 at 51 on 2026-07-27, versus 44 for DeepSeek V4 Pro and MiniMax M3 and 43 for Kimi K2.6. ^[1] The differences are more interesting than the ranking.

Scroll sideways for more →

Sources: the two model cards. ^[3][21] Comparable per-model specification tables are not published for every rival; DeepSeek and MiniMax disclose less about their 2026 flagships.

Against Kimi K2.7-Code, which landed on Hugging Face on June 12, 2026, GLM-5.2 trades expert count and multimodality for context length and license simplicity: Kimi's model is a trillion-parameter, 384-expert MoE with a vision encoder and a 256K window under a modified license, while GLM-5.2 is a denser-activating 40B-active model with four times the context, no image input, and plain MIT terms. ^[21] Against DeepSeek V4 Pro, the contrast CAISI drew is sharper on security than capability: DeepSeek's model was hijacked in 15.1 percent of scenarios by public attacks where GLM-5.2 was never hijacked, and it complied with every malicious cyber request under jailbreak. ^[9] Kimi K3 has since been scored on the same Artificial Analysis index, at 57 against GLM-5.2's 51, but Moonshot had not published its weights as of 2026-07-27, so GLM-5.2 remains the strongest downloadable model on that ranking. Qwen3.6 is not a later arrival at all: Alibaba published it in April 2026, two months before GLM-5.2, and the Qwen model Z.ai actually benchmarks against on its model card is the newer Qwen3.7-Max. ^[27][29][3]

Nathan Lambert, writing at Interconnects, called GLM-5.2 "the open weight model that feels right in coding harnesses as a general agent," and put the open-to-closed gap at roughly 6.8 months, counting the 204 days from Claude Opus 4.5 (November 24, 2025) to GLM-5.2. The UK AI Security Institute, testing GLM-5.2 and DeepSeek V4 Pro, reported on the same day as CAISI that "recent open weight models lag frontier closed models' cyber capabilities by 4 to 7 months," narrower than the 6 to 10 months it measured through most of 2025. CAISI's own report makes the comparison qualitatively, putting GLM-5.2's cyber capability level with Opus 4.6 from four months earlier, and gives no month range. ^[12][24][9] Because scores shift with the harness and the model version, these rankings reflect a mid-2026 snapshot rather than a fixed ordering.

How much does GLM-5.2 cost and where can you use it?

Z.ai's first-party API is priced at $1.40 per million input tokens, $0.26 per million cached input tokens, and $4.40 per million output tokens, the same rate card as GLM-5.1 and above GLM-5's $1.00 / $0.20 / $3.20. Z.ai lists cached-input storage as free on a limited-time basis. ^[17][1] The model also ships through third-party inference providers. OpenRouter listed more than 25 of them on 2026-07-27, at prices that mostly undercut Z.ai. Its default route was about $0.81 per million input and $2.53 per million output tokens, and the cheapest endpoints reached $0.76 and $2.42, but those serve FP4 or FP8 quantizations at a truncated context of 262,144 tokens or less rather than the full 1M window. Z.ai's own OpenRouter endpoint sits at the list price of $1.40 and $4.40, as do Fireworks, Together and Cloudflare, while DeepInfra is at $0.93 and $3.00. At launch there were only nine providers, "almost all of which are charging $1.40/million for input and $4.40/million for output." ^[7][11]

For subscription users, Z.ai sells the GLM Coding Plan, a flat-fee bundle that routes GLM-5.2, GLM-5-Turbo, and GLM-4.7 into third-party agents including Claude Code, Cline, and OpenCode. Z.ai's documentation lists three tiers starting at $18 per month with quotas of roughly 80, 400, and 1,600 prompts per five hours and 400, 2,000, and 8,000 per week, and states that the included quota is worth about 15 to 30 times the subscription fee at API rates. ^[18] GLM-5.2 and GLM-5-Turbo draw down that quota faster than the older models: Z.ai's blog set the multiplier at 3 times during peak hours (14:00 to 18:00 UTC+8) and 2 times off-peak, with a promotion billing off-peak use at 1 times through September 30, 2026. ^[2][18]

Z.ai gave ZCode, its own coding harness, a public launch on July 2, 2026, roughly two weeks after the weights, though its GLM-5.2 announcement post had already offered ZCode to Coding Plan subscribers in June. ZCode is marketed as the official harness for GLM-5.2 and is sold as part of the same subscription, though its documentation supports bring-your-own-key access to other vendors' models. Z.ai's own materials claim integration with more than 20 coding tools. Zixuan Li, the company's head of global operations, told the South China Morning Post that ZCode was built by "standing on the shoulders of an incredible open developer community." ^[8][20]

Can you run GLM-5.2 locally?

Only with serious hardware, or serious compression. The BF16 checkpoint is about 1.5 TB on disk, which Simon Willison described as "a 753B parameter, 1.51TB monster." ^[11] Unsloth published dynamic GGUF quantizations that mix bit depths per layer, bringing the 1-bit build to 223 GB and the 2-bit build to 245 GB, against 810 GB at 8-bit. Unsloth reports roughly 76 percent top-1 token agreement at dynamic 1-bit and about 82 percent at 2-bit, and describes 4-bit and 5-bit as near-lossless. ^[23] At those sizes even the smallest build needs a multi-GPU workstation or a high-memory Mac with headroom above 223 GB, not a single consumer card.

NVIDIA published an NVFP4 quantization for Blackwell-class hardware, and the model is packaged for Ollama, LM Studio, and Jan through the llama.cpp ecosystem. Hugging Face listed 119 quantized derivatives of GLM-5.2 as of 2026-07-27. ^[22] Z.ai has not released a smaller distilled variant of GLM-5.2 in the manner of GLM-4.5-Air; the only official checkpoints are the full BF16 model and its FP8 counterpart. ^[13]

How widely has GLM-5.2 been adopted?

Adoption has been fast by open-weight standards. As of 2026-07-27 the FP8 checkpoint reported about 3.1 million downloads in the previous month and the BF16 checkpoint about 1.0 million, with roughly 4,500 likes on the main repository. ^[16][3] The derivative ecosystem is where most of the volume sits: NVIDIA's NVFP4 build and Unsloth's dynamic GGUFs are the two most-downloaded quantizations, and 119 quantized variants were listed in total. ^[22]

The model is served by Z.ai's own API, by ZCode, and by third-party inference providers and gateways including OpenRouter and DeepInfra. Its position on agent-facing leaderboards drew attention as much as its benchmark scores: Lambert noted it competing with Anthropic's models on Arena's agent and design leaderboards, an unusual result for an open-weight release. ^[12][7]

What are GLM-5.2's limitations?

Three recur across independent write-ups.

The first is verbosity. Artificial Analysis measured about 43,000 output tokens per Index task against 26,000 for GLM-5.1 and 24,000 for MiniMax M3, which nearly doubles GLM-5.1's cost per task despite an identical price per token. Relaying the same Artificial Analysis measurement, Simon Willison wrote that they "did however find it to be quite token-hungry." Z.ai's answer is the selectable reasoning effort levels, which trade latency and token count against accuracy; Unsloth's deployment guide describes non-thinking, high, and max modes. ^[1][11][23]

The second is the absence of image input. GLM-5.2 takes text only, which rules it out of multimodal agent work and makes its strong showing on frontend and web-development evaluations somewhat surprising. Willison noted exactly that, and Lambert reported that the Claude Code harness trying to send images to the model "would brick Fireworks API for the session." ^[11][12]

The third is the gap between vendor and independent coding results. CAISI's contamination-resistant PortBench put GLM-5.2 at 41.7 percent, well behind models Z.ai's own tables place it alongside, which CAISI set directly against Z.ai's claim of "capability roughly positioned between Claude Opus 4.7 and Claude Opus 4.8." ^[9] For a model whose main selling point is agentic software engineering, that is the open question about GLM-5.2 rather than a footnote to it.

References

1. ^"GLM-5.2 is the new leading open weights model on the Artificial Analysis Intelligence Index." Artificial Analysis, 2026-06. artificialanalysis.ai/...alysis-intelligence-index
2. ^"GLM-5.2: Built for Long-Horizon Tasks." Z.ai / Zhipu AI (Hugging Face blog), 2026-06-17. huggingface.co/...glm-52-blog
3. ^"zai-org/GLM-5.2 model card." Hugging Face, 2026. huggingface.co/...GLM-5.2
4. ^"zai-org/GLM-5.2 config.json." Hugging Face, 2026. huggingface.co/...config.json
5. ^"GLM-5.2." Z.ai developer documentation (Chinese), 2026. docs.bigmodel.cn/...glm-5.2
6. ^"Z.ai's open-weights GLM-5.2 beats GPT-5.5 on multiple long-horizon coding benchmarks for 1/6th the cost." VentureBeat, 2026-06. venturebeat.com/...g-benchmarks-for-1-6th-the-cost
7. ^"GLM 5.2 API pricing and benchmarks." OpenRouter, accessed 2026-07-27. openrouter.ai/...glm-5.2
8. ^"Zhipu AI releases harness for GLM-5.2 model as Chinese firm takes aim at Anthropic." South China Morning Post, 2026-07-02. scmp.com/...model-chinese-firm-takes-aim-anthropic
9. ^"Assessment of Z.ai's GLM-5.2." Center for AI Standards and Innovation, NIST, 2026-07. nist.gov/...Assessment%20of%20Z.ai's%20GLM-5.2.pdf
10. ^"CAISI Assessment of Z.ai's GLM-5.2." NIST, 2026-07-17. nist.gov/...caisi-assessment-zais-glm-52
11. ^Willison, Simon. "GLM-5.2 is probably the most powerful text-only open weights LLM." simonwillison.net, 2026-06-17. simonwillison.net/...glm-52
12. ^Lambert, Nathan. "GLM-5.2 is the step change for open agents." Interconnects, 2026-06. interconnects.ai/...52-is-the-step-change-for-open
13. ^"zai-org/GLM-5 repository." GitHub, 2026. github.com/...GLM-5
14. ^"LICENSE (MIT), zai-org/GLM-5.2." Hugging Face, 2026. huggingface.co/...LICENSE
15. ^"LICENSE (Apache License 2.0), zai-org/GLM-5." GitHub, 2026. raw.githubusercontent.com/...LICENSE
16. ^"zai-org/GLM-5.2-FP8 model card." Hugging Face, accessed 2026-07-27. huggingface.co/...GLM-5.2-FP8
17. ^"Pricing." Z.AI Developer Documentation, accessed 2026-07-27. docs.z.ai/...pricing
18. ^"GLM Coding Plan overview." Z.AI Developer Documentation, accessed 2026-07-27. docs.z.ai/...overview
19. ^GLM-5 Team. "GLM-5: from Vibe Coding to Agentic Engineering." arXiv:2602.15763, 2026-02. arxiv.org/...2602.15763
20. ^"ZCode: Official Harness for GLM-5.2." Z.ai, accessed 2026-07-27. zcode.z.ai/en
21. ^"moonshotai/Kimi-K2.7-Code model card." Hugging Face, 2026-06. huggingface.co/...Kimi-K2.7-Code
22. ^"Quantizations of zai-org/GLM-5.2." Hugging Face, accessed 2026-07-27. huggingface.co/models
23. ^"GLM-5.2: How to Run Locally." Unsloth Documentation, accessed 2026-07-27. unsloth.ai/...glm-5.2
24. ^"How Far Behind the Frontier are Leading Open Weight Models on Cyber?" UK AI Security Institute, 2026-07-17. aisi.gov.uk/...leading-open-weight-models-on-cyber
25. ^Bai, Yushi, et al. "IndexCache: Accelerating Sparse Attention via Cross-Layer Index Reuse." arXiv:2603.12201, 2026-03-12. arxiv.org/...2603.12201
26. ^"Terminal-Bench 2.1 leaderboard." Terminal-Bench, accessed 2026-07-27. tbench.ai/...2.1
27. ^"Kimi K3 achieves #3 in the Artificial Analysis Intelligence Index." Artificial Analysis, 2026-07-17. artificialanalysis.ai/...e-to-opus-4-8-and-gpt-5-5
28. ^"GLM-5.2 model page." Artificial Analysis, accessed 2026-07-27. artificialanalysis.ai/...glm-5-2
29. ^"Qwen/Qwen3.6-35B-A3B model page." Hugging Face, 2026-04-15. huggingface.co/...Qwen3.6-35B-A3B