GLM-5.3 Flash: The Model That Was Hiding in Plain Sight
For six days, one of the most capable models on OpenCode and OpenRouter had no name. It was just called Ox Alpha. Then Z.ai revealed it: GLM-5.3 Flash, a 320B MoE model with a 1M token context, native multimodal input, MIT-licensed weights — and a price tag that makes you read it twice.
On August 20, 2026, an anonymous model appeared on OpenCode and OpenRouter with no explanation. It called itself Ox Alpha. It offered a one-million-token context window, image and video input, and a 131K output limit — entirely free, for a week, while quietly serving 100 trillion tokens a day. People noticed. Then on August 26, Z.ai (formerly Zhipu AI) pulled back the curtain: Ox Alpha was GLM-5.3 Flash, and the free run had been a product demo for something much bigger. Not just a new model — proof that Chinese AI infrastructure can handle frontier-scale traffic without a single Nvidia chip in the loop.
What It Actually Is
GLM-5.3 Flash is a 320-billion-parameter mixture-of-experts model. It activates 18 billion of those parameters per token — a design that lets it draw on a massive knowledge base while keeping per-token compute low. The architecture mixes 34 linear-attention layers with 11 sparse-attention layers, which cuts KV-cache costs significantly compared to a standard transformer. That hybrid is why the context window can sit at 1,048,576 tokens without the economics falling apart. It is the first natively multimodal model in the GLM-5 family. Previous GLM models handled text. This one was built from the ground up to take images, video, and files — not as an add-on, but at the architecture level. The weights are MIT-licensed. You can download them, fine-tune them, and ship them in a commercial product without restriction. The full repository is around 328 GB in FP8 or BF16 format on Hugging Face and ModelScope. That rules out a casual laptop run, but the licence makes serious self-hosting straightforward.
Where It Came From
GLM-5.3 Flash shares its base model with GLM-5.2. The gains are entirely from post-training: more executable environments, more varied long-horizon tasks, and more reinforcement-learning compute applied on top of the existing foundation. That framing matters. Z.ai is not claiming a new pretraining run. They are showing how far a frontier lab can move a model after pretraining ends — and the answer, going by the benchmarks, is quite far. The model was pretrained on 30 trillion multimodal tokens. That is a vendor disclosure rather than an independently audited number, but it gives you a sense of the scale. The full data recipe for the Flash-specific post-training is not public.
The Benchmarks
On DeepSWE v1.1, GLM-5.3 Flash scored 63.4. Its predecessor GLM-5.2 scored 46.2 on the same benchmark. That is a 17-point jump on a hard coding evaluation — from post-training alone. On the Artificial Analysis Intelligence Index, it scored 57 points. For context, that puts it alongside Claude Opus 4.8 on that particular measure. On Z.ai's GDPval-AA v2 agent benchmark (1773 Elo), it sits ahead of DeepSeek V4 Vision Exp (1675), Claude Opus 4.8 (1582), GPT-5.6 Terra (1571), and Gemini 3.7 Flash (1527). On Terminal-Bench 2.1 — the public leaderboard for sustained multi-step terminal agents — the independent standing at reporting time was 84.3, below Kimi K3 (88.3) and DeepSeek V4 Pro (87.9). That is still serious territory, and agent harnesses and scaffolds vary enough that the exact ranking shifts depending on the setup. The CyberSecurity angle is unusual: Z.ai reports an 84.5% score on CyberGym, up from 77.2% for GLM-5.2. They say the improvement in vulnerability-discovery skill 'exceeded expectations.' The post-training data included security-specific environments, so the direction was intentional — but the degree surprised even the team. The honest read: GLM-5.3 Flash belongs in the serious agent tier. Z.ai's charts are vendor-reported and the harness details matter. The numbers are credible enough to run your own evals — they are not numbers you dismiss.
The Price
API access through Z.ai costs $0.15 per million input tokens and $0.50 per million output tokens. Cached input drops to $0.03 per million. A launch promotion that runs through September 9, 2026 halves all of that — $0.075 in, $0.25 out. To put those numbers in context: Claude Opus 4.8 costs $5 per million input tokens and $25 per million output. GPT-5.6 Terra costs $2 in and $12 out. Z.ai describes GLM-5.3 Flash as roughly one-tenth the cost of comparable options, and on the output side that comparison largely holds. For a typical coding task — 16,000 input tokens, 1,080 output tokens — the bill at list price is about $0.003. At the promo rate, $0.0015. Fifty thousand such tasks would cost under $150 at list price, around $73 at the promo rate. On Multos, GLM-5.3 Flash is on the Lite tier via Novita — available on every paid plan, Lite and above. You do not need to manage API keys, set up endpoints, or track billing separately. It is just in the model picker.
The Chinese Chip Story
This is the part of the launch that caught the most attention outside the AI community. During the Ox Alpha preview week, Z.ai served 100 trillion tokens per day across clusters of domestically manufactured Chinese AI accelerators — not Nvidia GPUs. They built a custom SGLang-based serving stack tuned to that hardware and applied quantization techniques specific to it. The result was a threefold improvement in throughput compared to a generic deployment approach. US export controls have blocked Chinese companies from buying Nvidia's most advanced chips. Most assumptions about Chinese AI capability have quietly included 'but they still need NVIDIA for this part.' GLM-5.3 Flash is Z.ai's pointed answer to that assumption: frontier inference, frontier scale, no NVIDIA required. To be precise: Z.ai confirmed the inference ran on Chinese chips. They have not published a full account of what the training hardware was. The inference achievement stands on its own regardless.
What You Actually Get on Multos
When you pick GLM-5.3 Flash on Multos — either manually from the dropdown or via Auto Router — here is what the experience looks like in practice. First, the context. 1M tokens means you can paste an entire codebase, a full conversation history, a long document and a set of instructions all in the same call. On Multos, Multos manages context compaction automatically, so you are not manually tracking token counts or worrying about what to trim. GLM-5.3 Flash's window is large enough that compaction triggers less often than it does with most models, which means fewer interruptions mid-task. Second, multimodal input works straight from the chat. If you attach a screenshot — a UI bug, a design mockup, an error screen — GLM-5.3 Flash processes it natively. You do not need to switch to a different model or describe the image in text first. This is the first Lite-tier model on Multos where that is true. Third, the response quality. GLM-5.3 Flash sits in the same intelligence tier as Claude Opus 4.8 on general reasoning (AA Intelligence Index: 57 each). On coding specifically — DeepSWE v1.1: 63.4 — it is noticeably stronger than GLM-5.2 and competitive with models that cost significantly more. For everyday Multos tasks — writing code, debugging, explaining a codebase, drafting — the output quality is firmly in the tier that used to cost Starter or Pro pricing. Fourth, speed. The model activates 18B parameters per token rather than the full 320B, which keeps inference fast. Novita's serving infrastructure is optimised for low latency, and Multos routes to Novita for this model. In practice, responses feel quick — comparable to other flash-tier models on the platform. Finally, cost. GLM-5.3 Flash is on Multos Lite — the $20/month plan. You are not paying for a Pro tier to access a frontier-quality model. It counts against your Lite token pool alongside DeepSeek V4 Flash and MiMo V2.5. If you are already on Lite and wondering whether the plan is good enough for real work, GLM-5.3 Flash is a meaningful part of the answer.
What to Use It For
The combination of 1M token context, native image and video input, and aggressive pricing opens up some genuinely useful scenarios: Large codebase analysis — send a full repository worth of context in a single call without chunking. The context window is claimed, not independently validated at scale, but it is long enough to handle most real projects whole. Visual debugging and UI work — attach a screenshot or screen recording and ask what is happening. This is native, not bolted on. High-volume pipelines — at $0.15/M input, classification, extraction, and routing tasks become very cheap to run at scale. Cache hits drop that further to $0.03/M. Cost-sensitive agent workflows — if you are running multi-step agents and the token bills are becoming a concern, GLM-5.3 Flash at these rates is worth a serious look alongside DeepSeek V4 Flash. Where it may not be the first choice: if you need the full OpenAI tool stack (hosted shell, computer use, MCP, apply patch), or if you are heavily reliant on prompt caching at DeepSeek rates (DeepSeek V4 Flash caches cheaper), or if you need a provider with a long track record in production.
How It Compares to What's Already on Multos Lite
Multos Lite has a few models at this price band: DeepSeek V4 Flash (Novita and Fireworks), MiMo V2.5 (Novita), and now GLM-5.3 Flash (Novita). DeepSeek V4 Flash is cheaper on cached input ($0.014/M vs $0.03/M) and output ($0.28/M vs $0.50/M). It's the lowest-cost option for pure text and code at volume. GLM-5.3 Flash beats it on the AI Intelligence Index (57 vs lower) and adds native multimodal input, which DeepSeek V4 Flash doesn't have. MiMo V2.5 is fast and cheap ($0.168/M input) and strong on SWE-Pro benchmarks (56.1%), but it's text-only and smaller in scale than GLM-5.3 Flash's 320B parameter pool. GLM-5.3 Flash is the first Lite-tier model on Multos with native image and video input. If your task involves visuals, it's currently the only option at that price point that handles them natively.
Ready to Get Started?
GLM-5.3 Flash is live on Multos Lite and above via Novita. Native image and video input, 1M token context, frontier-tier intelligence at flash-tier pricing. Select it from the model dropdown or let Auto Router pick it — it's available from day one on the $20/month plan.
Start Building Free