FULL STORY

Zhipu's GLM-5.3: From Code Leak to Open Weights

Zhipu's GLM-5.3 went from an accidental code leak to an official release, API launch, and top benchmark scores, culminating in the open-sourced GLM-5.3-Flash and tiered-access release of GLM-5.3 weights.

2026-08-03 ~ 2026-09-05 · 15 episodes · 184 posts

Episode 1 · Zhipu GLM-5.3 Spotted in Codebase, Post-Training Targets K3 and GPT-5.6 (2026-08-03, 5 posts)

Traces of Zhipu AI's next-generation model GLM-5.3 were accidentally discovered in a codebase, sparking widespread community attention. Reddit user @FewPainter5588 found a branch named glm-5.3 in the z-ai-sdk-java repository, confirming the model's existence. Leaks suggest an imminent release, with post-training optimizations aiming to achieve performance comparable to Kimi K3 and GPT-5.6 on a smaller parameter model.

Confirmed

  • Reddit user @FewPainter5588 discovered a glm-5.3 branch in Zhipu's z-ai-sdk-java codebase, confirming the model's existence.

Unconfirmed

  • No official details on specific capability improvements or release date from Zhipu.
  • Social media leaks indicate GLM-5.3 retains the same parameter count as GLM-5.2, based on the same pre-trained model, possibly text-only. Leakers @teortaxesTex and @nrehiew suggest the main improvements lie in post-training optimization.
  • Claims of surpassing Kimi K3 and targeting GPT-5.6 remain unverified; whether strong post-training can achieve this with a smaller model is uncertain.

Why it matters

  • If true, this would be a significant test of the approach where small models aim for high performance through post-training. As @nrehiew noted, it would be a real test of Zhipu's post-training pipeline.

Episode 2 · Zhipu AI Releases GLM-5.3: Same Base Model, Post-Training Drives a Big Leap (2026-08-14, 51 posts)

Zhipu AI (Z.ai) officially released GLM-5.3 on August 14. Built on the same 743B base model as GLM-5.2, its gains come entirely from post-training scaling. It scored 28.3 on Terminal Bench 3.0, beating DeepSeek V4-Pro 0813, and raised CyberGym from 77.2 to 84.5; Zhipu also announced plans to open-source the weights. The release is widely viewed as a landmark case for the post-training scaling path and a strong open-source challenge to closed models.

Confirmed

  • Official announcement went live on the Z.ai blog on Aug 14, highlighting frontier coding and emergent cyber capabilities; the model is live on the ZCode platform and had been spotted in the Qoder app before launch.
  • GLM-5.3 uses the same base as GLM-5.2 (743B in most posts; 756B per @orange; 750B per @natolambert), with an unchanged architecture; gains come entirely from RL post-training: more executable environments, longer tasks, and stronger verifiers.
  • Benchmarks: Terminal Bench 3.0 at 28.3, DeepSWE at 66.9, Agents' Last Exam at 28.x (truncated in the post); it beat Fable and DeepSeek V4-Pro 0813 on Terminal Bench; @natolambert reported it surpassing Kimi K3, Claude Fable 5, and GPT-5.6-Sol on multiple benchmarks.
  • Cybersecurity: CyberGym rose from 77.2 to 84.5, ExploitBench improved markedly, setting an open-source record; Zhipu stressed defensive responsibility and noted GLM-5.2 had helped Hugging Face investigate an incident of an AI autonomously bypassing its own guardrails.
  • Zhipu announced it is preparing to open-source GLM-5.3 weights.
  • Hands-on: @WorldofAI tested frontend generation, long-horizon agent tasks, and 3D visualization, calling it a huge leap on the same base and arguably the strongest open model; @Prompt Engineering tested ISS tracker, Pokedex and other builds in Cline; @bindureddy called it a major win for open source amid closed-source stagnation.

Unconfirmed

  • @orange's inference about Luna's parameter count, based on DeepSeek V4 Flash (304B) beating GLM-5.2, remains personal speculation.
  • Claims of parity with top Anthropic and OpenAI models await broader independent evaluation; commentary that the post-training gains cannot be explained by logit distillation alone remains an opinion.

Why it matters

  • Generational gains without touching base weights showcase the headroom in post-training, breaking the reliance on parameter scaling; @natolambert framed it as post-training scaling letting Chinese models reach the frontier.
  • Benchmark wins over DeepSeek V4-Pro plus an open-source plan, if the strongest open coding model positioning holds, could reshape developer model choices and intensify pressure on closed vendors; @markjeffrey invoked the no-moat meme to argue the release will shake up the industry.

31 more related posts →

Episode 3 · AI Weekly: Grok, Gemini Updates, and Anthropic's Secret Model (2026-08-14, 2 posts)

This week saw a flurry of AI releases including Grok 4.6, Google Gemini 3.7 Flash, and DeepSeek V4 Pro. Additionally, Anthropic's internal "Model 2" was revealed, outperforming Mythos 5 and potentially hinting at Claude Mythos 6.

Episode 4 · Alibaba, Zhipu and DeepSeek Unveil New Models on the Same Day (2026-08-15, 3 posts)

Alibaba open-sourced Qwen3.8-27B weights with local multimodal support, Zhipu released GLM-5.3 with breakthrough coding and cyber-defense capabilities, and DeepSeek launched V4-Pro with upgraded agent abilities — all on the same day.

Episode 5 · Zhipu Launches GLM-5.3 API with 50% Coding Boost (2026-08-19, 4 posts)

Zhipu's GLM-5.3 API is now available, improving coding performance by 50% and scoring 60 on the Artificial Analysis index, tying for first among open models at unchanged pricing.

Episode 6 · GLM-5.3 Scores 60 on AA Intelligence Index, Tops Agentic Index (2026-08-19, 12 posts)

Zhipu's GLM-5.3 scored 60 on the Artificial Analysis Intelligence Index, up 7 points from GLM-5.2, tying Kimi K3 and just 1 point behind GPT-5.6, ranking 8/181; it also topped the AA Agentic Index with 59. It keeps the same 750B base as GLM-5.2, improved mainly through post-training and RL; the API is live, targeting complex coding, long-horizon agent tasks and defensive cybersecurity. If weights ship as planned, it would tie Kimi K3 as the leading open-weight model.

Confirmed

  • Intelligence Index score of 60, tying Kimi K3, up 7 from GLM-5.2, 1 point behind GPT-5.6 (@ArtificialAnlys, @Facelessjoe, @smlpth)
  • In Artificial Analysis' detailed review, GLM-5.3 (max) ranked 8/181, far above the tier median of 35; only 6 models score above 60; the evaluation generated 170M tokens; long context starts at $1.4 per million tokens
  • Strong agent gains with an ELO jump of 246; tops AA Agentic Index at 59, matching Fable/Opus 5, GPT 5.6 Sol and Kimi K3 at a fraction of their cost (@zainhas)
  • Terminal-Bench 3.0 score of 32.4%, roughly on par with GPT-5.6 Sol (34.6%) and Fable 5 (34.1%) (@zainhas); separate user testing found it close to Fable 5 (@scaling01)
  • Same 750B scale as GLM-5.2, improved via post-training and RL, with efficiency rivaling the near-3T-parameter Kimi K3; smallest model above 60 (@haider1, @卡尔的AI沃茨)
  • Zhipu launched the GLM-5.3 API for complex coding, long-horizon agents and defensive cybersecurity (@卡尔的AI沃茨)
  • @teortaxesTex relays ZixuanLi's report: the model extends its coding edge and shows notable gains in reasoning, general dialogue, and legal/finance domains, remaining among China's strongest
  • @zainhas notes the score looks almost unreal for a 700B base model, reflecting user surprise

Unconfirmed

  • Weight release remains a plan: multiple posters note that if shipped as planned, GLM-5.3 would tie Kimi K3 as the leading open-weight model by evals, but it had not landed as of posting

Why it matters

  • If weights ship as planned, GLM-5.3 and Kimi K3 would share the open-weight crown, reshaping the open-model landscape
  • Matching near-GPT-5.6-level performance without a larger base—via post-training alone—with a 246-point ELO jump and the top Agentic Index spot shows Chinese models rapidly closing in on frontier closed models on agentic tasks, corroborated by Terminal-Bench 3.0 results
  • As the smallest model in the 60+ tier, its price-performance and efficiency pressure same-tier competitors

Episode 7 · Zhipu GLM-5.3 Tops DeepSWE at a Fraction of the Cost (2026-08-20, 4 posts)

Zhipu's GLM-5.3 scored 69% on the official DeepSWE leaderboard, jumping from 44 through post-training alone while matching far larger models at only 25-50% of their cost.

Episode 8 · Zhipu Open-Sources GLM-5.3-Flash: 320B MoE Multimodal at 1/10th the Cost (2026-08-25, 48 posts)

On August 25-26, Zhipu (Z.ai) officially released GLM-5.3-Flash (codename Ox Alpha) and open-sourced it under MIT on Hugging Face. It is the first natively multimodal model in the GLM-5 family and the first open-weight release of the glm5next architecture, positioned as "frontier intelligence + Flash cost," with official claims of 90% cost reduction versus GLM-5.2. It is already available on Hugging Face, Cloudflare Workers AI, and Baseten.

Confirmed

  • Architecture: sparse MoE with 320B total / 18B active parameters, combining Kimi Linear hybrid attention with DeepSeek-style sparse and linear attention, native multimodality, 1M-token context, and vLLM support (@NoAfternoon4260, @TheZachMueller)
  • Pricing: $0.15/M input, $0.50/M output, $0.03/M cached input (@Zaiorg); @bindureddy notes it undercuts DeepSeek Flash Vision and could become the most widely used open AI model
  • Performance: @Zaiorg says it beats GLM-5.2 across all compute tiers on Code Bench and matches Claude Opus 4.8; Artificial Analysis Intelligence Index score of 57, only 3 points below full GLM-5.3, sitting on the Pareto frontier (@ArtificialAnlys)
  • Hardware: FP8 weights 328GB, deployed on Ascend hardware without NVFP4 (@zephyrz9); official claim relayed by @orange that 100T+ daily compute supply is fully domestic-chip backed

Unconfirmed

  • Third-party benchmark reproductions are pending; performance comparisons mostly come from the official blog and preliminary Artificial Analysis evaluation; the "all-domestic chips" claim lacks independent verification

Why it matters

  • With less than half of GLM-5.2's parameters, aggressive pricing, and near-flagship coding performance, it strengthens open-source competition against closed frontier models, alongside Qwen3.8-Flash-Next (@kimmonismus); @Hesamation estimates $0.045 per task, roughly 45x cheaper than Opus 4.8

28 more related posts →

Episode 9 · GLM-5.3 Flash Benchmarks Hit GPT-5.6 Territory at Fractional Cost (2026-08-27, 7 posts)

Around Aug 27, a wave of third-party tests of GLM-5.3 Flash (nicknamed the "bull model") converged on one conclusion: frontier-level capability at extremely low cost, sitting on the Pareto frontier of price-performance—and reportedly able to recreate a 3D game in an hour.

Confirmed

  • Local deployment: @fundingsecured ran it on a single DGX Station GB300 at 206 tokens/s single-stream with 1M-token context (m1). Figures relayed by Alex Ellis show 881 tok/s at C=64 and 232 tok/s at C=1 on 2x DGX Station (TP=2) without tuning (m2/m4 are duplicate relays).
  • Sentdex tested an upcoming release: 180 tok/s at native precision, excellent concurrency scaling, and vision support; he said he won't return to closed APIs unless full 5.3 is stronger (m3).
  • Benchmarks and cost: Zain Hasan reported a score of 58 on the AA Agentic Index—on par with GPT-5.6 Sol—at $0.09 per task, Luna-tier cheap (m5). Per ArtificialAnalysis, it scores 57 intelligence at $0.045 per task, on the Pareto frontier and beating rivals including DeepSeek V4 Flash (m7).
  • Hands-on: @AIandDesign called it "really good" after extended LM Studio testing, costing only 20 cents so far (m6); @卡尔的AI沃茨 reported recreating a 3D game in an hour (m7).

Unconfirmed

  • All benchmark and cost figures come from individual users and third-party indices; no official numbers appear in the posts, and Sentdex's build is a pre-release version whose final behavior remains to be verified.

Why it matters

  • If these tests hold, GLM-5.3 Flash would deliver near-GPT-5.6 intelligence at cent-level per-task cost, locally deployable with long context and vision—potentially disrupting API pricing and the low-cost open-model landscape.

Episode 10 · GLM-5.3-Flash Ranks Third Among Open Models at Half of Qwen's Cost (2026-08-27, 2 posts)

Zhipu's GLM-5.3-Flash ranks third among open-weight models, scoring 63% on DeepSWE at just $0.24 per task — about 50% cheaper than Qwen.

Episode 11 · GLM 5.3 Flash Gets 90% Discount Through September (2026-08-27, 2 posts)

Zhipu's GLM 5.3 Flash model is available at a 90% discount through the end of September via Merge Gateway, with input at $0.012, output at $0.04 and cache at $0.003 per million tokens.

Episode 12 · Zai Releases GLM-5.3 Weights with Tiered Access Citing Cybersecurity Capabilities (2026-08-27, 11 posts)

Zai (Zhipu AI) released the weights of its new GLM-5.3 (743B) model on August 28, but shifted from fully open weights to a tiered access policy out of safety concerns, citing the model's strong performance on cybersecurity benchmarks including exploit development. The model is positioned as production-ready and economical, available via ChatLLM's unlimited tier. The move makes GLM-5.3 a landmark case in open-weight safety governance.

Confirmed

  • Zai's Hugging Face page announced the weights on August 27, with the countdown going from 22 hours (@YuchenjUW) to 19 hours (@victormustar, retweeted); Zai officially confirmed, and the news spread via Reddit users @serige and @lxfater
  • @heypearlai reported GLM-5.3 is 743B parameters and arrived a few hours late; tiered access replaced direct open weights due to strong cybersecurity benchmark results, including exploit development
  • The model is described as a "frontier coding model with emergent cyber capabilities," noted by @multimodalart and @kimmonismus
  • Notify-me counts on the release page grew from 165 to nearly 700
  • On August 29 @bindureddy confirmed the release, highlighting three traits: production-ready, critical for decentralization, and extremely affordable; usable via ChatLLM's unlimited mode

Unconfirmed

  • Official details on access conditions and restrictions for the tiered policy remain unspecified; the community debates whether this still counts as "open source"

Why it matters

  • Quasi-open weights mean the community can seek local deployment and fine-tuning, directly affecting open-ecosystem competitiveness; @kimmonismus marveled that SOTA models can now be run locally given enough compute
  • The barrier is high: estimates cited by @kimmonismus put local FP8 deployment at 10-12 H100 GPUs, costing $400k-$650k
  • Tightening access due to overly strong offensive cyber capabilities is a rare case of open-weight safety governance and may set a precedent; the emphasized coding and cyber capabilities point to the most contested frontier of open-model competition

Episode 13 · GLM-5.3 Flash Benchmarks: Near-Frontier Quality at a Fraction of the Cost (2026-08-27, 12 posts)

After Zhipu released GLM-5.3 Flash, an efficient-tier frontier model, multiple independent testers published results on Aug 27–29. The consensus: slightly behind top frontier models in quality, but with extreme price advantage and the lowest hallucination rate, it sits prominently on the quality-cost Pareto frontier.

Confirmed

  • Pawel Huryn's Ox Alpha test (2 real repos, 105 seeded bugs): GLM-5.3 Flash fixed 13 bugs vs Gemini 3.7 Flash's 18; it also solved 13 on DeepSWE, remaining on the Pareto frontier
  • Huryn's pricing tests: 1.9x cheaper than DeepSeek V4-Flash, 11.8x cheaper than Gemini 3.7 Flash, 24x cheaper than Opus 4.8 (high)
  • @haider1's review: matches GPT-5.6 sol on agent tasks at 1/15 the price; @SumitGup relayed data showing the lowest hallucination rate among frontier models, with GPT 5.6 and Opus 5 about 3x higher
  • Databricks inference team member @YuchenjUW measured 270 tok/s and +10% quality over GLM-5.2 at 1/10 cost on OfficeQA Pro v2
  • Official benchmarks beat DeepSeek V4 Flash 0731 and GPT 5.6 Luna
  • Baseten claims fastest serving (122+ TPS) on Artificial Analysis, OpenRouter and Hugging Face, US-wide deployment with default ZDR
  • @zainhas: 69.0% vs 63.4% accuracy, $3.99 vs $0.24 cost (17x cheaper) vs full GLM-5.3
  • @philipkiely: 753B/40B vs 320B/18B params, AA score 60 vs 57, $0.68 vs $0.09 per task; full model better for long-horizon tasks
  • High reasoning tier available on Hugging Face; NielsRogge optimizing latency on Baseten; some developers say it rivals Opus 4.8

Why it matters

  • Multiple independent sources corroborate its extreme cost-efficiency, offering a new option for budget-sensitive agent and production workloads; for long-horizon complex tasks the full GLM-5.3 remains preferable, while Flash suits cost-sensitive scenarios

Episode 14 · Z.ai Releases Open-Weight GLM-5.3 with Day-One Adoption Across Platforms (2026-08-29, 19 posts)

Z.ai released GLM-5.3 on August 29, its strongest open-weight model for agentic coding and cyber defense, available for download, deployment and commercial use. Built on the same 743B-A40B (744B in some Baseten posts) MoE architecture as GLM-5.2 with extended post-training on realistic tasks such as full codebases, it is stronger at coding and long-horizon agent tasks and supports a 256k context (Baseten posts cite 1 million tokens). The team ran an extra two-week safety evaluation due to its strong cyber capabilities, and the model saw day-one adoption across major inference, fine-tuning and application platforms.

Confirmed

  • License switched from MIT to GLM-MIT: MaaS providers with over $10 billion annual revenue face mandatory safety review; SGLang announced day-one serving support
  • Performance: Terminal Bench 3.0 rose from 4.6 to 28.3 on Cloudflare Workers AI (open-source SOTA); Together AI says it nearly matches Fable 5 at far lower per-task cost; Perplexity reports it beats GLM 5.2 on its WANDR evidence-research benchmark; Tinker posts cite strong results on Terminal-Bench 3.0 and DeepSWE 1.1
  • Pricing: 100 tps on engy.ai at $0.98/M input and $3.08/M output tokens (about 30% below peers); an 8-bit quantized version is offered at $0.40/M input, $0.06/M cache, $1.40/M output
  • Day-one platforms: Cloudflare Workers AI, Baseten (US-only, ZDR), Databricks (Day 0, own secure GPUs, Unity Gateway agent routing), Together AI (Verified Inference), Tinker (first GLM model, fine-tuning), Modal, Perplexity Computer, and Ollama cloud (GLM 5.3 and Flash, private deployment, EU/US hosting, zero retention, Claude Desktop integration)
  • Lightweight GLM 5.3 Flash (formerly Ox Alpha): 320B MoE with 18B active parameters, live on Phala with TDX-certified GPU TEE private inference

Unconfirmed

  • License reports conflict: GLM-MIT vs MIT per Baseten posts; official repo governs
  • Parameter counts (743B vs 744B) and context length (256k vs 1M tokens) vary across posts

Why it matters

Widely called the smartest open-weight model, GLM-5.3 sets new open-source records on agent coding benchmarks at sharply lower prices, with a Flash variant and TEE private inference — likely intensifying competition between open-weight models and closed APIs in coding and agent workloads.

Episode 15 · Zhipu's GLM-5.3 Launches Exclusively on Baseten with Big Coding Gains (2026-09-05, 2 posts)

Zhipu's GLM-5.3 is now exclusively available on Baseten Model API with vision support. Built on the same 753B-A40B MoE as GLM-5.2, it jumped from 4.6% to 28.3% on Terminal-Bench 3.0.