--- title: GLM-5.3 review: Zai's new open model tops my coding and... description: Zai's newly released GLM-5.3 model tops a personal coding and agentic benchmark suite (Kingbench 3), scoring 73/80 (91.25%), beating Fable 5, Qwen 3.8 Max, and... canonical: https://daily.dev/posts/glm-5-3-review-zai-s-new-open-model-tops-my-coding-and-agentic-benchmark-61vfxby1s twitter:card: summary_large_image twitter:site: @dailydotdev og:type: website og:site_name: daily.dev og:title: GLM-5.3 review: Zai's new open model tops my coding and agentic benchmark | daily.dev og:description: Zai's newly released GLM-5.3 model tops a personal coding and agentic benchmark suite (Kingbench 3), scoring 73/80 (91.25%), beating Fable 5, Qwen 3.8 Max, and... og:url: https://daily.dev/posts/glm-5-3-review-zai-s-new-open-model-tops-my-coding-and-agentic-benchmark-61vfxby1s og:image: https://api.daily.dev/og/posts/61vfxBY1S.png og:image:alt: GLM-5.3 review: Zai's new open model tops my coding and agentic benchmark og:image:width: 1200 og:image:height: 630 og:locale: en --- > ## Documentation Index > Fetch the complete documentation index at: https://daily.dev/llms.txt > Use this file to discover all available pages before exploring further. # GLM-5.3 review: Zai's new open model tops my coding and agentic benchmark **[Collections](https://daily.dev/sources/collections)** · 2 min read · 0 upvotes · 0 comments ## Summary Zai's newly released GLM-5.3 model tops a personal coding and agentic benchmark suite (Kingbench 3), scoring 73/80 (91.25%), beating Fable 5, Qwen 3.8 Max, and Opus 4.8/5. Despite sharing the same parameter count and architecture as GLM-5.2, post-training improvements pushed the score up from 75%. The release is pitched around security specialization, including code auditing and vulnerability discovery, validated with outside security teams, alongside an 'open-source shield initiative' that keeps defensive capabilities open while gating high-risk misuse features. Coding performance was strong on both frontend and backend tasks, and the model cracked a previously unsolved 3D generation prompt (a contact-lens case) plus a hard wristwatch generation task. Math and agentic results were also solid. ## Content Z.ai (Zhipu) released GLM-5.3 this week, and the story isn't really about coding benchmarks anymore, even though that's ostensibly what the model is for. It's about how good this thing turned out to be at finding security vulnerabilities, and how fast that capability showed up during training. ## What actually shipped GLM-5.3 is a post-training-only upgrade. Same base model as GLM-5.2, same parameter count, same architecture. Z.ai says the entire performance jump came from scaling up reinforcement learning, specifically about ten times more long-horizon task environments than before. No new pretraining run. And the jump is not subtle: - Terminal-Bench 3.0: 4.6 → 28.3 - DeepSWE v1.1: 46.2 → 66.9 - CyberGym: 77.2% → 84.5% That CyberGym number is the one turning heads. At 84.5%, GLM-5.3 edges out Anthropic's Mythos 5 (83.8%) and OpenAI's GPT-5.6 Sol (83.6%) on real-world vulnerability discovery. A Chinese open-weight model, built on a base that's reportedly smaller than its Western rivals, beating them at finding bugs in code. Z.ai's own line on this: "Scaling post-training is all we did for GLM-5.3." But it's not a clean sweep. On ExploitBench, which tests actually building working exploits rather than just spotting flaws, GLM-5.3 scores 54.4% against 78.0% for Mythos 5 and 76.5% for GPT-5.6 Sol. Same story on ExploitGym: 105 completed tasks in two hours versus Mythos 5's 181. So the model is good at finding problems and considerably less good at weaponizing them. Whether that gap closes in the next version is the question everyone's quietly asking. ## The vulnerability hunt Working with Chinese security teams, Zhipu ran GLM-5.3 against 269 real-world projects and came back with 2,436 vulnerabilities. That includes 107 critical and 990 high-severity findings, some reportedly dating back roughly 40 years. Fifty-three have been disclosed publicly; the rest sit under embargo. Z.ai is tracking all of it in a public Security Disclosure Ledger. One detail that's hard to shake: the model apparently found a "potentially serious vulnerability" in Cursor, the AI coding tool itself. There's something a little uncomfortable about an AI model auditing the tools built to contain AI models. ## Where the gains actually came from There's genuine disagreement about the mechanism here, and I don't think it's fully settled. Z.ai insists this is pure RL scaling on the existing base model, not distillation from something bigger. Some observers aren't convinced. One theory floating around is that ## Questions this post answers ### How does GLM-5.3 compare to GLM-5.2 and other models like Opus 4.8 on coding benchmarks? GLM-5.3 scored 73 out of 80 (91.25%) on the Kingbench 3 coding, 3D simulation, math, and agentic test suite, the highest score recorded on that benchmark, ahead of Fable 5, Qwen 3.8 Max, and both Opus 4.8 and Opus 5. GLM-5.2 had only reached 75% two months earlier despite sharing the same parameter count and architecture, indicating the gain came from post-training changes rather than scale. _Developers picking a coding model can track fresh benchmark comparisons like this on daily.dev._ ### What new capability does GLM-5.3 add compared to previous GLM models? GLM-5.3 adds a dedicated specialization in security analysis, including code auditing and vulnerability discovery, marketed under the tagline 'Built to Code. Ready for Cyber Defense.' Zai says this was validated with actual security teams rather than only internal benchmarks, and pairs it with an 'open-source shield initiative' that keeps defensive security features open while gating capabilities that could enable high-risk misuse. _Teams evaluating AI tools for secure coding workflows can follow releases like this on daily.dev._ ## Community take How the wider developer community reacted, aggregated from 2 discussions and 179 comments across hackernews, x (as of 2026-08-21). **TL;DR:** Most of the discussion isn't about GLM-5.3 itself but about speculation over a separate mystery model ('Ox Alpha'), with people debating whether it's distilled from a bigger teacher, a stealth release from another lab, or just a smaller Flash-tier model; direct commentary on GLM-5.3 is sparse and mixed. GLM-5.3: Frontier Coding with Emergent Cyber Capabilities: Commenters are impressed that a post-training-only update on the same base model closes much of the gap to closed frontier models at a fraction of the size and cost, while a large side-thread debates whether US labs' cybersecurity guardrails and access restrictions are justified or just gatekeeping/fear-mongering. **Sentiment:** 31% positive · 45% mixed · 24% skeptical **The case for** - A few people who tried it reported being genuinely impressed, especially with a related free model available in Open Code. - Many see it as remarkable that gains come purely from post-training on the same base model, suggesting post-training is now the key lever for progress. - One commenter argued real RL in verifiable execution environments, not just distillation, likely explains the strong agentic/SWE performance. - Several note it's impressively close to closed frontier models despite being a much smaller parameter count. - Being free/open and cheap to run is seen as a major practical advantage over closed frontier offerings. **The pushback** - Several argued the compute required to sustain a free, high-performing model like this seems implausible for smaller labs, casting doubt on claims about its origins. - Some argue closed frontier models (Fable/Sol/Mythos) still lead, especially at the high end of exploitation-chain benchmarks. - One benchmark tester said it seemed like a small, coding-only Flash-tier model that couldn't handle serious agentic tasks and wasn't close to top-tier models. - Concerns that quantizing the model locally lobotomizes capability, making local deployment a tradeoff. - One user said they weren't impressed when testing it in plain conversation, though they hadn't tried coding or agentic tasks. **By community** - hackernews (heated): Excitement about the benchmark gains and openness is real, but the thread is dominated by a heated, sprawling debate over whether closed-lab cybersecurity guardrails are legitimate safety measures or self-serving gatekeeping. - x (mixed): Replies mostly speculate about the identity and training method of a related mystery model rather than assessing GLM-5.3 directly, with opinions split between distillation, stealth-lab, and small-model theories. **Hottest debate:** Whether the performance jump comes from genuine RL scaling or from distillation off a larger undisclosed teacher model. **Open questions** - Is the mystery 'Ox Alpha' model actually related to GLM-5.3, and if so what is its parameter scale? - Will the promised open-source weight release in two weeks actually happen without further restrictions? - Does the releasing organization have enough compute to sustain offering such a model for free? - How will consumer hardware (e.g., upcoming Apple silicon) realistically handle running a model this large at usable speed and quality? - Is scaling post-training/RL indefinitely sustainable for continued capability gains, or will it plateau? **Highlights** > @scaling01 Based on our benchmark results, this is definitely a small Flash-tier model, not a large one. It is clearly trained purely on coding and is just too small to handle serious agentic tasks. That is totally fine for a fast code model, but it is definitely not Opus level. Not even > — [foodtruckbench on x](https://x.com/foodtruckbench/status/2090761367171338657) > @scaling01 Distillation being the boring explanation is exactly why I believe it, big teacher quietly squeezed into a small student is the least glamorous story anyone could tell about a jump like that. > — [\_ocooper on x](https://x.com/_ocooper/status/2090699002614366635) > @scaling01 Distillation can transfer general style and priors, but hitting frontier-level SWE and agentic task success almost always requires real RL in verifiable execution environments. Even if there’s a giant teacher model behind it, the feedback loop is what actually seals the gap. > — [seastartech on x](https://x.com/seastartech/status/2090688370716664041) > @scaling01 Does Zai has that much compute ? Everyone is confused for that part > — [AdamOptimizers on x](https://x.com/AdamOptimizers/status/2090729425063338010) > @scaling01 I can't say I was really impressed. Only tried it in plain conversation/deep thought prompting, not coding, design, or anything agentic. Since Muse 1.2 was also available free in Open Code, I tried it out as well. Was genuinely impressed. > — [Pharmguy6 on x](https://x.com/Pharmguy6/status/2090712440921063774) **Source threads** - [hackernews](https://news.ycombinator.com/item?id=49294997) · 300 points · 126 comments - [x](https://x.com/scaling01/status/2090662468833976582) · 0 points · 53 comments --- Tags: [#ai](https://daily.dev/tags/ai), [#llm](https://daily.dev/tags/llm) [View this post on daily.dev](https://daily.dev/posts/glm-5-3-review-zai-s-new-open-model-tops-my-coding-and-agentic-benchmark-61vfxby1s) ```json {"@context":"https://schema.org","@graph":[{"@type":"Organization","@id":"https://daily.dev/#organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180},"sameAs":["https://twitter.com/dailydotdev","https://github.com/dailydotdev","https://www.linkedin.com/company/daily-dev-ltd"]},{"@type":"WebSite","@id":"https://daily.dev/#website","url":"https://daily.dev","name":"daily.dev","publisher":{"@id":"https://daily.dev/#organization"},"potentialAction":{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https://daily.dev/search?q={search_term_string}"},"query-input":"required name=search_term_string"}}]} {"@context":"https://schema.org","@type":"TechArticle","headline":"GLM-5.3 review: Zai's new open model tops my coding and agentic benchmark","url":"https://daily.dev/posts/glm-5-3-review-zai-s-new-open-model-tops-my-coding-and-agentic-benchmark-61vfxby1s","mainEntityOfPage":{"@type":"WebPage","@id":"https://daily.dev/posts/glm-5-3-review-zai-s-new-open-model-tops-my-coding-and-agentic-benchmark-61vfxby1s"},"datePublished":"2026-08-14T05:39:55.219Z","dateModified":"2026-08-21T13:50:29.795Z","description":"Zai's newly released GLM-5.3 model tops a personal coding and agentic benchmark suite (Kingbench 3), scoring 73/80 (91.25%), beating Fable 5, Qwen 3.8 Max, and...","isAccessibleForFree":true,"articleSection":"Collections","inLanguage":"en","publisher":{"@type":"Organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180}},"author":{"@type":"Organization","name":"Collections","logo":"https://media.daily.dev/image/upload/s--fk_6ycEi--/f_auto,q_auto/v1780996001/logos/collections?_a=BAMAMiWQ0","url":"https://daily.dev/sources/collections"},"commentCount":0,"discussionUrl":"https://daily.dev/posts/glm-5-3-review-zai-s-new-open-model-tops-my-coding-and-agentic-benchmark-61vfxby1s","interactionStatistic":[{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":0},{"@type":"InteractionCounter","interactionType":{"@type":"CommentAction"},"userInteractionCount":0}],"keywords":"ai,llm","timeRequired":"PT2M"} {"@context":"https://schema.org","@type":"BreadcrumbList","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https://daily.dev"},{"@type":"ListItem","position":2,"name":"Collections","item":"https://daily.dev/sources/collections"},{"@type":"ListItem","position":3,"name":"GLM-5.3 review: Zai's new open model tops my coding and agentic benchmark"}]} {"@context":"https://schema.org","@type":"FAQPage","@id":"https://daily.dev/posts/glm-5-3-review-zai-s-new-open-model-tops-my-coding-and-agentic-benchmark-61vfxby1s#faq","mainEntity":[{"@type":"Question","name":"How does GLM-5.3 compare to GLM-5.2 and other models like Opus 4.8 on coding benchmarks?","acceptedAnswer":{"@type":"Answer","text":"GLM-5.3 scored 73 out of 80 (91.25%) on the Kingbench 3 coding, 3D simulation, math, and agentic test suite, the highest score recorded on that benchmark, ahead of Fable 5, Qwen 3.8 Max, and both Opus 4.8 and Opus 5. GLM-5.2 had only reached 75% two months earlier despite sharing the same parameter count and architecture, indicating the gain came from post-training changes rather than scale. Developers picking a coding model can track fresh benchmark comparisons like this on daily.dev."}},{"@type":"Question","name":"What new capability does GLM-5.3 add compared to previous GLM models?","acceptedAnswer":{"@type":"Answer","text":"GLM-5.3 adds a dedicated specialization in security analysis, including code auditing and vulnerability discovery, marketed under the tagline 'Built to Code. Ready for Cyber Defense.' Zai says this was validated with actual security teams rather than only internal benchmarks, and pairs it with an 'open-source shield initiative' that keeps defensive security features open while gating capabilities that could enable high-risk misuse. Teams evaluating AI tools for secure coding workflows can follow releases like this on daily.dev."}}]} ```