GLM-5.3-Flash on Run BiOS

GLM-5.3-Flash is the first natively multimodal model in the GLM-5 series from Z.ai. With 320B total parameters and just 18B active per token (320B-A18B), Z.ai reports that it outperforms GLM-5.2 across benchmarks and real-world workloads at one-tenth the price, while approaching Claude Opus 4.8 on coding and agentic benchmarks.

GLM-5.3-Flash starts from a newly trained base model, with its architecture and training recipe redesigned around capability and efficiency. For the first time in the GLM series, it introduces a hybrid architecture combining sparse and linear attention, sharply reducing long-context serving costs while preserving precise long-context capabilities. The model also adopts Manifold-Constrained Hyper-Connections (mHC) to further improve scaling efficiency. Together with Z.ai's latest 30T-token multimodal pre-training corpus, these changes enable GLM-5.3-Flash to deliver more intelligence with less compute.

Architecture

GLM-5.3-Flash is a mixture-of-experts (MoE) model with a hybrid sparse/linear attention design and a native vision encoder. Published model configuration:

• Scale: 320B total parameters with 18B active per token (320B-A18B); BF16 weights.
• Context: a maximum position embedding of 1,048,576 tokens (1M context).
• Language model: 45 hidden layers (the first 3 dense, the remainder sparse MoE), hidden size 4096, 288 routed experts with 8 routed experts active per token plus 1 shared expert, MoE intermediate size 2048, dense-layer intermediate size 12288, 64 attention heads, MLA-style attention with Q LoRA rank 1536 and KV LoRA rank 512, sigmoid expert scoring, vocabulary size 154,880.
• Sparse attention indexer: 32 indexer heads (head dimension 128) selecting the top 2048 positions per query, with indexer sharing extended to the MTP iteration.
• Hyper-connections: Manifold-Constrained Hyper-Connections (mHC) with an expansion factor of 4.
• Speculative decoding: one multi-token-prediction (MTP) layer.
• Vision encoder (native multimodal): a 24-layer vision transformer, hidden size 1024, patch size 14, base image size 448 with 2× spatial merge, and a temporal patch size of 2 for video input; images and videos are projected into the language model through a 4096-dimensional interface. The model accepts text, image, and video inputs.

Thinking and Reasoning Effort

Per the official model card:

• 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 (or if set to any other value). To use low or high, pass them explicitly.
• For benchmark and leaderboard reproduction, Z.ai recommends keeping the default max.
• In the model's chat template, clear_thinking defaults to false; Z.ai recommends explicitly passing clear_thinking=true for chat scenarios.

Performance

Z.ai's published claim for this release: GLM-5.3-Flash outperforms GLM-5.2 across benchmarks and real-world workloads at one-tenth the price, while approaching Claude Opus 4.8 on coding and agentic benchmarks. The full benchmark chart is published in the official model card; evaluation methodology notes published alongside it include:

• HLE w/ tools (full set): sampling temperature=1.0, top_p=0.95, maximum generation length 163,840 tokens, maximum context length 300,000 tokens with a context management strategy; GPT-5.6-luna (medium) as the judge model.
• NL2Repo: temperature=1.0, top_p=1.0, max_new_tokens=64k under 1M context, with rule-based and LLM-based judgement to prevent malicious behaviors (e.g., unauthorized pip or curl operations).
• DeepSWE: the mini-swe-agent harness with temperature=0.95, top_p=1.0, timeout=6h, and 400K context.
• Terminal-Bench 2.1: evaluated in Claude Code 2.1.207 with temperature=1.0, top_p=1, max_new_tokens=65536, and a 6-hour timeout.
• Toolathlon Verified: results from the official evaluation service, pass@1 averaged over 3 independent runs.
• AutomationBench: evaluated on v1.0.6, incorporating the fix for the null-type handling issue.
• BabyVision: temperature=1.0, top_p=0.95, maximum context length 164K tokens, input images resized so the shorter side is at least 1.5K pixels.

License

GLM-5.3-Flash is released under the MIT license.

On Run BiOS

On the Run BiOS serverless catalog this model is served as glm-5.3-flash with its full 1,048,576-token context window, function calling, structured outputs, multimodal input on supported routes, adjustable reasoning effort (exposed as a low / high / max ladder, matching the model's own three levels; the model's own default is max), and prompt caching for repeated prefixes.