GLM-5.3-Flash

GLM-5.3-Flash is Zhipu's (Z.ai) first natively multimodal GLM-5 series model, released August 26, 2026 and confirmed to be the "Ox Alpha" model that appeared stealth on OpenRouter days earlier. It is a 320B-parameter mixture-of-experts model with only 18B active parameters per token, a 1M-token context window, and MIT-licensed open weights. At $0.15 per million input tokens it is positioned as a cheap, long-context, open-weight option for agents.

Model Specifications

Key Features

• Natively multimodal: unlike earlier GLM vision adaptations, this model was pre-trained from scratch on a 30-trillion-token multimodal corpus, so it reads image and video input natively.
• Hybrid sparse + linear attention: the GLM series' first blend of sparse attention and linear attention layers cuts attention compute by about 3x and KV cache size by about 4.4x versus GLM-5.3.
• 1M context, cheap to serve: the smaller KV cache is what makes long-context self-hosting practical; a 320B MoE still needs serious memory to serve, but far less than full-attention would.
• MIT open weights: weights are downloadable from Hugging Face under the MIT license, so self-hosting and commercial use are both permitted.
• Always-on thinking: reasoning cannot be disabled, so every response spends reasoning tokens; Z.ai recommends temperature 1.0 and top_p 0.95 with maximum reasoning effort.
• Coding Plan integration: included in GLM Coding Plan with a 3x quota increase and 50% off during off-peak hours.

Use Cases

Who Should Use This Tool?

• Agent builders running long reasoning loops where repeated context dominates cost, since the $0.03 cached-input rate makes re-sent context nearly free.
• Multimodal pipelines that need image and video understanding in one open-weight model.
• Self-hosters with enough GPU memory to serve a 320B MoE, where the 4.4x smaller KV cache makes 1M-context feasible.
• Cost-sensitive teams comparing against closed frontier models that cost several times more per task.

Problems It Solves

1. Long-context serving cost: full attention scales quadratically and forces a large KV cache; the hybrid design reduces both.
2. Multimodal fragmentation: one model replaces separate image and video encoders.
3. Vendor lock-in: MIT open weights let a team leave a hosted provider while keeping the same model.

Pricing

Advantages & Unique Selling Points

Compared to Competitors:

1. Versus GLM-5.3: a stronger and far cheaper long-context serving profile, at the cost of a smaller active parameter count.
2. Versus DeepSeek-V4-Flash and Kimi-K3: Zhipu claims the lowest attention compute among the baselines it measured, while acknowledging its KV cache stays slightly larger than some rivals.
3. Versus closed frontier models: roughly an order of magnitude cheaper per task for comparable long-context agentic work.

What Makes It Stand Out:

• First natively multimodal GLM, pre-trained, not adapted.
• Runs entirely on Chinese AI chips, per Z.ai.
• The same model shipped anonymously as "Ox Alpha", an unusual no-training-carve-out stealth preview.

User Reviews

Community attention was intense. Patrick Collison called the stealth "Ox Alpha" preview "very impressive", which sent a wave of developers to test it. Fingerprint hunters used tokenizer probes to declare it "99% certain" to be a Zhipu GLM model before the company said a word. Benchmarkers noted it hit strong coding scores, though community testing on serving platforms flagged it as slower than some flash peers under certain deployment conditions.

Getting Started

Quick Start Guide

1. Via API: call model ID glm-5.3-flash on either Z.ai's open platform or Zhipu's bigmodel platform, using the standard OpenAI-compatible chat format.
2. Via OpenRouter: the stealth/ox-alpha listing gives way to a first-party GLM-5.3-Flash listing; the free window closed around August 27, so expect standard pricing.
3. Self-host: download weights from zai-org/GLM-5.3-Flash on Hugging Face, then serve with your runtime of choice; use a quantized build if VRAM is tight.

Integration

• OpenAI-compatible chat, function calling, streaming, and JSON structured output.
• Image input via URL or Base64.
• Hugging Face Transformers, llama.cpp, and Unsloth (GGUF) for local serving.

Frequently Asked Questions

Is GLM-5.3-Flash the same as GLM-5.3?

No. GLM-5.3-Flash is a newly trained base model, not a distillation of GLM-5.3, and it is natively multimodal where GLM-5.3 is text-only.

Can I disable thinking?

No. Thinking is always on. Every response spends reasoning tokens, so plan token budgets accordingly.

Is it really open source?

Weights are available under the MIT license, which permits commercial and self-hosted use. As with any open-weight release, the training data is not disclosed.

What is 'Ox Alpha'?

Ox Alpha was the cloak name for GLM-5.3-Flash, confirmed by Z.ai on August 26, 2026 after it appeared anonymously on OpenRouter and leaned on independent community usage before branding.

Alternatives

• GLM-5.3: the text-only coding flagship with stronger cyber capability but fewer active savings.
• Qwen3.8-Flash-Next: another August 26 open-weight cost-efficiency release, 125B total / 6B active.
• DeepSeek V4 Flash: a strong open model for cost-sensitive agentic deployments.

Tips & Best Practices

1. Budget for thinking tokens: reasoning can't be turned off, so estimate per-task cost on repeated reasoning, not single turns.
2. Use cached input: agentic loops resend system prompts and context; at $0.03/M cached, context becomes nearly free after the first pass.
3. Quantize for single-rig serving: a 320B MoE with 18B active still needs serious VRAM; pick a quant that fits or offload to RAM.

Conclusion

GLM-5.3-Flash is Zhipu's bet on cheap, long-context, natively multimodal open weights, wrapped around an attention redesign that makes 1M-context serving economical. For agent builders and multimodal pipelines that care about per-task cost rather than peak benchmarks, it is one of the strongest value options in the open-weight space right now. Start at the Z.ai announcement or the Hugging Face model card.