Z.ai shipped GLM-5.2 weights to Hugging Face on June 16, 2026. Three days earlier, Coding Plan subscribers had already been running it. The lab dropped weights first and benchmark numbers second, compressing the usual launch-narrative window into a 72-hour discovery period. That release sequence-quiet ship, then numbers, then hype-turned out to be smart, because by the time anyone could push back on the claims, the model had already been stress-tested inside real coding workflows.
The result is an open-weight model that is genuinely hard to dismiss. With 744 billion parameters, a 1 million token context window, and an MIT license, GLM-5.2 performs close to Anthropic's Claude Opus 4.8 and OpenAI's GPT-5.5 on coding and agentic benchmarks while costing roughly one-sixth the price.
What GLM-5.2 actually scores on coding benchmarks
The confirmed benchmark scores are: SWE-bench Pro 62.1, Terminal-Bench 2.1 at 81.0 (the first open-weight model to exceed 80%), FrontierSWE at 74.4% (within 1% of Claude Opus 4.8), and MCP-Atlas at 77.0. Those numbers are not vendor-reported-the lab notably published no benchmark table at launch- Z.ai has not published a full technical report, and the headline numbers come from third-party evaluations (Artificial Analysis), VentureBeat reporting, and community reviews.
The scores held up under independent scrutiny, which is what matters. GLM-5.2 reached 51 on the Artificial Analysis Intelligence Index-the new top open-weight score and 4th overall-and on Code Arena: Frontend, the Max tier hit 1,595, surpassing every Claude Opus variant.
The jump from its predecessor is worth noting. Compared to GLM-5.1: the context window grew from 200K to 1 million tokens, Terminal-Bench rose from 63.5 to 81.0, DeepSWE jumped from 18.0 to 46.2, and High and Max thinking levels were added. That is not a minor version bump in practice.
Two caveats deserve the same space as the scores. First, at least one prominent commentator calls it bench-maxxed, while others note GLM-5.1 scored 0.0% on at least one benchmark GLM-5.2 now performs well on. Second-and this one is concrete- GLM-5.2 does not support image understanding, which limits it on visual analysis workflows. If your coding agents need to read screenshots, diagrams, or UI mockups, this is a real constraint, not a footnote.
The MIT license and why it changes the cost math
The license is the real story here, not the benchmarks. An open-weight model is one whose trained parameters are publicly available for download, allowing organizations to run it on their own hardware, fine-tune it on their own data, and deploy it without routing requests through the original developer's servers. GLM-5.2 goes further: it ships under the MIT license, which allows nearly every use case, including commercial work.
Apache 2.0 and MIT-level licensing terms mean a downstream organization can download, fine-tune, and redistribute the model commercially without seeking the developer's permission or triggering a legal review. For teams under data-residency requirements-healthcare, finance, government-that matters more than any benchmark score.
The API pricing is competitive even if you never touch the weights. Z.ai's direct API is $1.40 per million input tokens and $4.40 per million output tokens; OpenRouter lists it as low as $0.95/$3.00 from some backing providers. There is also a GLM Coding Plan with tiered subscriptions starting at $12.60 per month for the Lite plan.
Z.ai claims this reduces per-token FLOPs by 2.9x at 1M context length, which matters because serving a 1M-context model at reasonable cost is where most comparable efforts have stumbled. This is a serving-efficiency claim, not a capability claim. It lowers the cost of long-context inference once the model is loaded-it does not reduce the memory to load it.
Beagle in action#eng-platform, 11:02amThe VRAM math that every "just self-host it" post skips
Here is the number that cuts through the hype: no single GPU can run GLM-5.2 on its own-even the most aggressive 1-bit quantization needs roughly 223 GB, more than any single card offers. The entry point for genuine GPU-speed inference is a 4× A100 80 GB instance.
The reason is specific to MoE architecture and is worth understanding. For a Mixture-of-Experts model, you cannot predict which experts a token will activate, so all expert weights must sit in VRAM even though only a fraction compute per token. That means weights memory tracks total parameters, not active parameters. GLM-5.2 fires ~40B parameters per token, but all 744B must be resident.
The quantization options and their practical implications:
| Precision | Weights memory | Minimum GPU config | Notes |
|---|---|---|---|
| BF16 (full) | ~1,488 GB | Multi-node cluster | Not practical for most teams |
| FP8 | ~744 GB | 8× H200 SXM | Reference serving config |
| INT4 / AWQ | ~372 GB | 4× H200 or 5× A100 80 GB | Near-lossless on most tasks |
| 2-bit GGUF (Unsloth) | ~239-245 GB | 4× RTX 3090 + 256 GB RAM | ~82% of full accuracy; 3-9 tok/s |
Unsloth's 2-bit dynamic GGUF compresses the model from 1.51 TB down to about 239 GB, so it can plausibly run on a 4× RTX 3090 rig with 192 GB system RAM, a 256 GB+ Mac Studio, or a high-RAM dual-socket workstation. Expect roughly 3 to 9 tokens per second on consumer hardware.
That speed matters. Three to nine tokens per second is workable for batch overnight jobs. It is not workable for an interactive coding agent answering in real time.
The production path is clearer: a single 8× H100 cluster (~$10/hr on most clouds) sustains hundreds of tokens per second of GLM-5.2 throughput; at the $4.40/M output price, breakeven against the Z.ai API arrives somewhere around 600M output tokens per month per cluster.
Using the Z.ai API at a typical 60/40 input/output token mix, the blended API price is roughly $2.60 per million tokens. If you're processing under that 600M token/month threshold, the API is almost certainly cheaper than owning the cluster.
One under-reported catch: maximum output length is host-dependent. Z.ai's docs say 128K, the Hugging Face card states up to 163,840 tokens for reasoning tasks, and OpenRouter enforces its own 32,768-token ceiling per response regardless of the model spec. If your agent tasks rely on long output chains, check the specific provider's ceiling before committing.
What is genuinely new here versus incremental
The honest read: GLM-5.2 is not a DeepSeek R1 moment for the whole field. It is a meaningful step for one specific workload-long-horizon agentic coding-where it competes within a few percentage points of the closed frontier at dramatically lower API cost.
The community reaction is comparable to the response to DeepSeek R1. That is not a comparison made lightly, and it has well exceeded the response to Kimi K2's release. But the scope is narrower than R1's. DeepSeek R1 disrupted the market broadly as a general-purpose model. GLM-5.2 is narrower in scope and aimed squarely at the specific workload where enterprise spending is accelerating fastest: agentic coding.
The architecture has two genuinely interesting pieces that are not just marketing. A 20% increase in speculative decoding acceptance length via an improved MTP layer points at inference speed as a real operational advantage. And the 1M-token context window is a five-fold expansion over GLM-5.1's 200K limit; Z.ai emphasizes that this window is "solid," maintaining coherent reasoning throughout rather than degrading at extended lengths as some nominally large-context models do.
The gap versus closed models on the hardest tasks is real. On FrontierSWE, GLM-5.2 is almost level with Opus 4.8 and beats GPT-5.5-but it falls behind on the hardest tests like SWE-Marathon, so there is still room to grow on very long tasks.
A teammate like Beagle evaluating whether to route a specific agent task to GLM-5.2 versus a closed API would weigh exactly this: task type, output volume, whether images are involved, and data-residency requirements. The model wins clearly on cost for high-volume code tasks with no visual input. It does not yet win on every hard agentic benchmark, and the no-image-input constraint is a real limit for full-stack engineering agents.
GLM-5.2 open-weight coding model: common questions
What is GLM-5.2 and who made it?
GLM-5.2 is a 744B-parameter Mixture-of-Experts open-weight language model released by Z.ai (formerly Zhipu AI) on June 13-16, 2026. It ships under an MIT license, scores 62.1 on SWE-bench Pro, and targets long-horizon agentic coding. API pricing starts at $1.40 input / $4.40 output per million tokens.
How does GLM-5.2 compare to GPT-5.5 and Claude Opus 4.8 on coding?
GLM-5.2 beats GPT-5.5 on SWE-bench Pro (62.1 vs 58.6) and FrontierSWE (74.4% vs 72.6%), at roughly one-sixth the output cost. It trails Claude Opus 4.8 on the hardest long-horizon tasks (SWE-marathon) and lacks image input, which Opus 4.8 supports.
Can you self-host GLM-5.2 on a single GPU?
No. Even the most compressed 1-bit quantization requires roughly 223 GB of combined memory-more than any single GPU offers. The practical entry point for production inference is a 4× A100 80 GB instance. Consumer hardware (4× RTX 3090 + 256 GB RAM) can run the 2-bit GGUF at 3-9 tokens per second, suitable for batch work only.
What is the GLM-5.2 self-hosting break-even point?
At Z.ai's API pricing of roughly $2.60 per million tokens blended (60/40 input/output), and a production 8× H100 cluster costing ~$10/hr, self-hosting breaks even at approximately 600 million output tokens per month per cluster. Below that volume, the managed API is almost always cheaper after accounting for engineering time.
Does GLM-5.2 support image and multimodal input?
No. GLM-5.2 accepts text and code input only-it has no image understanding. This is the clearest capability gap versus Claude Opus 4.8 and GPT-5.5. If your agent workflows involve reading diagrams, UI screenshots, or visual assets, GLM-5.2 is not a direct substitute.