FULL STORY

Zhipu Open-Sources GLM-5.3-Flash: From Launch to Hands-On Benchmarks

Zhipu released and open-sourced GLM-5.3-Flash under MIT with 90% lower costs. Hands-on benchmarks soon showed performance rivaling GPT-5.6.

2026-08-25 ~ 2026-08-27 · 2 episodes · 50 posts

Episode 1 · Zhipu Open-Sources GLM-5.3-Flash: 320B Multimodal MoE at ~90% Lower Cost (2026-08-25, 48 posts)

Zhipu officially released and open-sourced GLM-5.3-Flash (codename Ox Alpha) on August 26. It is the first natively multimodal model in the GLM-5 series and the first open-weights release of the glm5next architecture. Published on Hugging Face under the MIT license, it positions itself as "frontier intelligence + Flash speed at Flash cost," with the company claiming the price is just one-tenth of GLM-5.2's.

Confirmed

  • Architecture & scale: A sparse MoE architecture with 320B total parameters and 18B activated parameters, combining Kimi Linear hybrid attention with DeepSeek-related techniques (sparse attention + linear attention); natively multimodal, supports a 1 million token context window, and already has vLLM support
  • Pricing: Official standard API pricing is $0.15 per 1M input tokens, $0.50 per 1M output tokens, and $0.03 for cached input
  • Performance: Officially, it significantly outperforms GLM-5.2 across all compute tiers on the Code Bench real-world coding benchmark and matches Claude Opus 4.8; it scores 57 on the Artificial Analysis intelligence index, only 3 points below the full GLM-5.3, sitting on the Pareto frontier alongside GLM-5.3
  • Hardware deployment: According to @zephyrz9 and @orange, the FP8 weights are roughly 328GB and the model is deployed on domestic chips such as Ascend; the company says its daily supply of over 100T of compute is fully backed by domestic chips

Not Yet Confirmed

  • Third-party replications of the benchmark data are still pending; the claimed performance comparisons (such as matching Opus 4.8) currently come mainly from the official blog and preliminary Artificial Analysis evaluations

Why It Matters

  • With less than half the parameters of GLM-5.2 yet surpassing it across benchmarks, paired with ultra-low pricing, observers like @bindureddy see potential for it to become the most widely used open-source model; @Hesamation estimates its per-task cost at about $0.045 — 15x cheaper than GLM-5.3 and roughly 45x cheaper than Opus 4.8
  • All inference compute is backed by domestic chips; if true, this has major implications for the industry's supply landscape and represents a rare "performance-cost" sweet spot in the open-source community

28 more related posts →

Episode 2 · GLM-5.3 Flash Matches GPT-5.6 at a Fraction of the Cost (2026-08-27, 2 posts)

Early tests show GLM-5.3 Flash scores 58 on the AA Agentic Index, matching GPT-5.6 Sol at a fraction of the cost. Users running it in LM Studio report strong performance, spending only 20 cents in extended testing.