GLM-5.3-Flash

GLM-5.3-Flash

Quick answer: GLM-5.3-Flash is Zhipu AI's (Z.ai) first natively multimodal model in the GLM-5 series — a 320B-total/18B-active-parameter MoE released under the MIT license with a 1,048,576-token (1M) context window. It scores 84.3% on Terminal-Bench 2.1, 63.4% on DeepSWE v1.1, and 1,773 Elo on GDPVal-AA v2, beating its own larger sibling GLM-5.2 across the board at roughly one-tenth the serving cost, while approaching Claude Opus 4.8 on coding and agentic benchmarks.

At a Glance

Where GLM-5.3-Flash leads

• Open weights under MIT — fully self-hostable, unlike the proprietary GLM-5.3 flagship
• Outperforms the much larger GLM-5.2 on every benchmark Zhipu reports (Terminal-Bench 2.1 81.0%→84.3%, DeepSWE 46.2%→63.4%, GDPVal-AA v2 1,514→1,773 Elo) at a fraction of the inference cost
• First natively multimodal GLM-5 model — hybrid sparse+linear attention architecture keeps long-context serving cheap while adding image/video understanding
• Native 1M-token context window (1,048,576 tokens)

Where it lags

• Smaller/cheaper "Flash" tier — the proprietary GLM-5.3 flagship (same generation, API-only) scores higher on every shared benchmark (e.g. Terminal-Bench 2.1 88.2% vs. 84.3%)
• Agents' Last Exam (26.3%) trails frontier closed models (Claude Opus 4.8 27.3%, GPT-5.6 Terra 27.0%, Gemini 3.7 Flash 28.0%) by a small but consistent margin
• Output is text-only — no image/audio generation

Best for: Teams that want open, self-hostable multimodal coding/agent capability close to the GLM-5.3/Claude Opus tier without the proprietary-API lock-in or the full flagship's serving cost.

What GLM-5.3-Flash Is

GLM-5.3-Flash is Zhipu AI's (Z.ai) cost-efficient, fully open-weight entry in the GLM-5.3 generation, trained from a newly built base model rather than reusing GLM-5.2's pretrain. It's the first model in the GLM-5 series designed as natively multimodal from the ground up, pairing a 320B-total/18B-active mixture-of-experts language backbone with a dedicated vision tower for image and video understanding.

The headline architectural change is a hybrid attention design: most layers use linear attention (Gated DeltaNet-style, "KDA"), with a smaller number of full/sparse ("DeepSeek sparse attention") layers interleaved throughout the 45-layer stack. This sharply reduces the cost of serving very long contexts while preserving retrieval-style long-context accuracy — a departure from the pure full-attention or pure-MoE approach used by earlier GLM releases. The model also adopts Manifold-Constrained Hyper-Connections (mHC), a scaling-efficiency technique layered on top of the standard MoE routing (288 routed experts + 1 shared expert, top-8 routing per token). Zhipu trained it on a fresh 30-trillion-token multimodal pre-training corpus.

Before Zhipu's identity was confirmed, this model briefly circulated as an anonymous "stealth" release under third-party branding, generating outsized attention on OpenRouter and community leaderboards purely on capability signal before anyone knew who built it — a reminder that raw benchmark performance travels faster than provenance in this market. Now that it's confirmed as GLM-5.3-Flash with a real org, MIT license, and full HF model card, it stands as one of the strongest fully open-weight multimodal releases of August 2026.

Specifications

Pricing

Open weights (MIT) — free to download and self-host; hosting/inference cost only. Also available hosted via the Z.ai API — check Z.ai's pricing page for current per-token rates (not independently verified here).

Context Window

GLM-5.3-Flash has a 1,048,576-token (1M) context window — enough to hold entire large codebases or hundreds of pages of documents in a single request. The hybrid linear/sparse attention design specifically targets keeping this scale of context affordable to serve, rather than treating it as a stretch-goal extension.

Public Benchmark Scores

Scores above are reported by Zhipu AI/third parties and shown for context. They are not Benchgen measurements. Figures depend on harness, tools, and effort settings (default reasoning_effort=max) — see the source for methodology.

Additional benchmark reported by Zhipu, not yet on Benchgen's leaderboard:

• AutomationBench v1.0.6: 48.8% — this version includes a bug fix (null-type handling, PR #13) that measurably raises scores versus Benchgen's existing AutomationBench entries for other models (e.g. DeepSeek V4 Pro 31.8%, Kimi K3 30.8%, Qwen3.8 Max 27.3% on the pre-fix version) — the same version-mismatch reason the sibling GLM-5.3 page also excludes this score. Not added pending version reconciliation.

GLM-5.3-Flash vs Alternatives

*GLM-5.2 Terminal-Bench 2.1/DeepSWE figures are from Zhipu's own published comparison (not yet separately verified as Benchgen leaderboard entries for GLM-5.2). †Agents' Last Exam shown as a percentage here for readability; Benchgen's leaderboard stores it as a 0–1 fraction.

Despite being the "Flash" (cost-efficient) tier, GLM-5.3-Flash beats its own larger, more expensive predecessor GLM-5.2 on every metric Zhipu reports, and lands within a few points of Kimi K3 and DeepSeek V4 Pro on coding/agentic benchmarks — while being the only model in this comparison that's both natively multimodal and fully open-weight (MIT).

Related Models & Benchmarks

• Compare: GLM-5.3 · GLM-5.2 · Kimi K3 · DeepSeek V4 Pro
• Benchmarks: TerminalBench 2.1 · DeepSWE · Agents' Last Exam · Humanity's Last Exam · GDPVal-AA v2

Frequently Asked Questions

Specs from Zhipu AI's official HF model card (huggingface.co/zai-org/GLM-5.3-Flash) and Benchgen research. Last updated 2026-08-27.