GLM-5.3-Flash

GLM-5.3-Flash was served at scale on Chinese AI chips with a high-bandwidth interconnect and optimized serving stack. Key optimizations include:

• Intra-node tensor parallelism for Linear Attention and LM head
• ReplaySSM
• W8A8 quantization
• Hybrid INT8/FP8/BF16 cache quantization
• Layer Split
• Encode-Prefill-Decode (EPD) disaggregated architecture

A 3x improvement in end-to-end serving performance was achieved compared to baseline, reaching hardware efficiency and per-token cost comparable to mainstream NVIDIA GPUs. Notably, a GLM-5.3-powered infrastructure agent assisted engineers in optimizing the serving stack.

GLM-5.3-Flash was pre-trained on a 30T-token multimodal corpus, the latest in the GLM series. The combination of optimized pre-training corpus and architectural improvements enables the model to produce more intelligence with less compute.

GLM-5.3-Flash delivers frontier-level intelligence at approximately 1/10th the cost of comparable models. On the Artificial Analysis Intelligence Index v4.1.1, it scores 57 at just $0.045 per task (discounted).

The model was tested anonymously as 'ox-alpha' on OpenCode and OpenRouter before release, becoming the most popular model of the week \u2014 with all traffic served on Chinese AI chips.

GLM-5.3-Flash supports controlling the thinking budget through the reasoning_effort parameter, which accepts three levels: low, high, and max. It defaults to max if not passed. For benchmark and leaderboard reproduction, keep the default max.

In the chat template, clear_thinking defaults to false. For chat scenarios, explicitly pass clear_thinking=true.

GLM-5.3-Flash is the first natively multimodal model in the GLM-5 series. It supports text and image inputs, with multimodal pre-training on a 30T-token corpus.

GLM-5.3-Flash is the first model in the GLM series to introduce a hybrid architecture combining sparse and linear attention. This sharply reduces long-context serving costs while preserving precise long-context capabilities. The model also adopts Manifold-Constrained Hyper-Connections (mHC) to improve scaling efficiency.\n\nKey architecture stats:\n- Total parameters: 320B\n- Active parameters: 18B (per token)\n- Layers: 45 (compared to 92 in GLM-4.5)\n- Context length: 1M tokens\n\nCompared to GLM-4.5 (355B total / 32B active / 92 layers), GLM-5.3-Flash nearly halves both activated parameters and layer count while delivering superior performance.