Update — August 28, 2026: GLM-5.3-Flash's day-one MIT weights are now a useful contrast point: the separate, larger GLM-5.3 (non-Flash, cyber-defense line) missed its own staged-release target. Z.ai's Hugging Face placeholder listed August 28 for GLM-5.3's open weights, and that date passed with nothing published. See GLM-5.3 open weights delayed — Z.ai misses its own Aug 28 target for the full timeline. Those open weights also let Unsloth ship a Dynamic 3-bit GGUF sized to run on 128GB of RAM.
Update — August 30, 2026: Open-weight GLM-5.3 placed third on Terminal-Bench 4.0, ahead of GPT-5.6 Sol on the terminal-agent leaderboard — GLM-5.3 Terminal-Bench 4.0 third place.
Update — August 30, 2026: The MIT license also made GLM-5.3-Flash a target for third-party refusal removal — OrcaRouter shipped an uncensored build with the edit baked directly into the model's native block-FP8 weight shards, not a LoRA adapter or a prompt trick: OrcaRouter's uncensored GLM-5.3-Flash, explained.
Update — August 27, 2026: The launch date matters for a second story — three Google Gemini team posts from August 22 were retroactively read as trolling this launch, four days before it happened. See the corrected Ox Alpha timing timeline, which also puts GLM-5.3-Flash and Gemini 3.7 Flash side by side on price, license, and vendor benchmarks.
Update — September 4, 2026: GLM-5.3-Flash is running a launch promo through September 9, 2026 at $0.075/M input and $0.25/M output — about 18x cheaper than flagship GLM-5.3's $1.40/$4.40 list price. After the promo ends it reverts to $0.15/$0.50 list price, roughly a 9x discount. Full pricing breakdown, plus a fact-check of a separate viral claim about Claude Fable 5.1 and SimpleBench, in GLM-5.3 Flash's price cut and the Fable 5.1 SimpleBench claim, checked.
On August 26, 2026 at 7:42 PM, Z.ai (@Zai_org) posted the reveal the community spent six days chasing:
Introducing GLM-5.3-Flash — Previously previewed as Ox Alpha, running entirely on Chinese AI chips.
The tweet hit 740K+ views within an hour. Weights landed on Hugging Face under MIT License. API docs went live at docs.z.ai/guides/llm/glm-5.3-flash. The stealth codename stealth/ox-alpha on OpenRouter was always a preview skin — this is the named product.
If you followed the Ox Alpha forensics timeline, the ending is almost anticlimactic: stack traces, error code 1214, and 30/30 tokenizer matches to GLM-5.3 were right. Bloomberg confirmed Zhipu that morning; Z.ai confirmed the SKU and architecture that evening.
TL;DR — GLM-5.3-Flash at a glance
| Question | Answer |
|---|---|
| Former codename | Ox Alpha (stealth/ox-alpha on OpenRouter) |
| Architecture | 320B total / 18B active (MoE-style sparse activation) |
| Context | 1M tokens, natively multimodal (text, image, video) |
| License | MIT — weights on Hugging Face |
| API price | $0.15/M in · $0.50/M out · $0.03/M cached |
| Stealth infra | Entire preview served on Chinese AI chips, per Z.ai |
| Inference stacks | SGLang, vLLM, TokenSpeed (local); ZCode, Chat, AutoClaw |
| vs GLM-5.2 | Large gains on DeepSWE (+17.2 pts) and AutomationBench (+22.6 pts) |
| vs Opus 4.8 | Leads GDPVal-AA v2; trails Terminal Bench 2.1 slightly |
| Coding Plan | 3× usable quota vs GLM-5.3 on same subscription tier |
From Ox Alpha stealth to named model
The launch closes a six-day arc explainx.ai tracked in three acts:
| Date | Event |
|---|---|
| Aug 20 | Ox Alpha appears on OpenRouter — free, 1M context, anonymous Stealth provider |
| Aug 21–22 | Community forensics tie it to Z.AI / GLM-5.3 (full evidence) |
| Aug 26 AM | Bloomberg: Zhipu confirms GLM lineage; weights promised that night |
| Aug 26 PM | Z.ai: names GLM-5.3-Flash, ships MIT weights, publishes benchmarks |
Z.ai's launch blog is explicit about the strategy: test anonymously on OpenCode and OpenRouter, become "the most popular model of the week," then ship under the GLM brand with open weights. Hunter Alpha and Healer Alpha → MiMo was the same playbook from a different lab.
What changes for builders now:
- Route to
z-ai/glm-5.3-flash(or Hugging Face locally) instead of guessing aboutstealth/ox-alphalongevity - MIT license — self-host, fine-tune, strip guardrails (with the usual responsibility caveats)
- Multimodal native — not a text-only model with bolt-on vision; Z.ai frames vision as part of the coding loop (render → verify → refine)
Specs and architecture
From Z.ai's technical post:
| Spec | GLM-5.3-Flash | GLM-5.2 (prior) |
|---|---|---|
| Total params | 320B | (prior gen larger activated count) |
| Active params | 18B | 32B (GLM-4.5 series comparison in blog) |
| Layers | 45 | 92 (GLM-4.5 series) |
| Context | 1M tokens | Shorter in prior GLM-5.2 API tier |
| Attention | Hybrid linear + sparse + IndexPool for long context | Prior architecture |
| Pre-training | 30T-token multimodal corpus | — |
Architectural headline: Manifold-Constrained Hyper-Connections (mHC) plus hybrid attention cut long-context serving cost — Z.ai claims 3.0× less attention compute and 4.4× smaller KV cache vs GLM-5.3 (non-Flash) on per-token metrics.
Benchmarks — coding and agentic
Z.ai's published evaluation table (August 26, 2026):
| Benchmark | GLM-5.3-Flash | GLM-5.2 | DeepSeek-V4-Vision-Exp | Claude Opus 4.8 | GPT-5.6 Terra | Gemini 3.7 Flash |
|---|---|---|---|---|---|---|
| Terminal Bench 2.1 | 84.3 | 81.0 | 83.9 | 85.0 | 87.4 | 85.8 |
| DeepSWE v1.1 | 63.4 | 46.2 | 59.3 | 58.0 | 69.6 | 65.3 |
| AutomationBench v1.0.6 | 48.8 | 26.2 | 38.8 | 41.0 | 37.2 | 52.3 |
| Agents' Last Exam | 26.3 | 20.4 | 27.3 | 27.0 | 28.0 | — |
| HLE w/ Tools | 55.3 | 54.7 | 55.1 | 57.9 | — | — |
| GDPVal-AA v2 | 1773 | 1504 | 1675 | 1582 | 1571 | 1527 |
| Toolathlon Verified | 78.4 | 59.9 | 75.9 | 76.2 | 74.9 | — |
explainx.ai read:
- Biggest jump vs GLM-5.2: AutomationBench (48.8 vs 26.2) and DeepSWE (63.4 vs 46.2) — aligns with Ox Alpha's agent traffic on OpenRouter (Claude Code, Hermes Agent)
- GDPVal-AA v2 leadership is the headline for enterprise agentic workflows — beats Opus, GPT-5.6 Terra, and Gemini 3.7 Flash
- Not a blanket frontier win: GPT-5.6 Terra still leads DeepSWE (69.6) and Terminal Bench (87.4); Gemini 3.7 Flash leads AutomationBench (52.3)
- Independent Ox Alpha DeepSWE claim (80%) from Ben Davis during stealth may reflect harness differences vs Z.ai's official 63.4 — always run your own eval
On Z.ai Code Bench v1.0 (Claude Code 2.1.207 harness), Flash beats GLM-5.2 at every effort level and at max effort scores 29.0 vs Opus 4.8's 29.5 — functionally tied on Z.ai's in-house coding eval.
Pricing and where to run it
API (standard)
| Tier | Price per 1M tokens |
|---|---|
| Input | $0.15 |
| Output | $0.50 |
| Cached input | $0.03 |
Compare to Claude / GPT frontier pricing — Flash-tier MoE at sub-dollar output is the China playbook in action (free preview → named open model).
Platforms (all live Aug 26)
| Surface | URL |
|---|---|
| Weights | huggingface.co/zai-org/GLM-5.3-Flash |
| API | docs.z.ai/guides/llm/glm-5.3-flash |
| Coding Plan | z.ai/subscribe — 3× quota vs GLM-5.3 |
| ZCode | zcode.z.ai — Browser Use + Computer Use |
| Chat | chat.z.ai |
| AutoClaw | autoclaw.z.ai |
| OpenRouter | Likely z-ai/glm-5.3-flash soon if not already — watch model page |
Local inference
# Example — verify current flags against Z.ai docs before production
# SGLang / vLLM support announced at launch; TokenSpeed also listed
Z.ai notes SGLang, vLLM, and TokenSpeed at launch; more frameworks incoming. For agent harness setup patterns, see how to run GLM-5.2 on coding agents — routing is the same, swap model ID.
Chinese AI chips — the infra story
The line that will age into geopolitics textbooks:
"All of this traffic served on Chinese AI chips."
Z.ai's blog describes:
- Large-scale cluster of domestic accelerators with high-bandwidth interconnect
- Custom SGLang inference engine — partially built with GLM-5.3 infrastructure agents (model optimizing its own serving stack)
- EPD disaggregation — Encode / Prefill / Decode worker pools at cluster scale
- 3× end-to-end serving improvement vs baseline on same hardware, comparable per-token cost to NVIDIA GPUs
For builders outside China, the practical takeaway is not nationalism — it is proof that a MoE Flash model can serve at OpenRouter scale without H100 clusters, which matters for inference cost deflation forecasts.
What people are asking
"Should I stop using stealth/ox-alpha on OpenRouter?"
Migrate to the named model when OpenRouter lists it — stealth previews can disappear or reprice without notice. Until the slug updates, stealth/ox-alpha may still work but treat it as deprecated.
"Is this the same as GLM-5.3 from August 14?"
No — different SKUs. GLM-5.3 (Aug 14) is the 743B-base cyber-defense line with staged open weights. GLM-5.3-Flash is a 320B-A18B multimodal Flash variant with immediate MIT release. Same family name, different products.
"MIT license — can I use it commercially?"
Yes, MIT is permissive. Standard caveats: verify the Hugging Face repo license file, run your own safety evals for regulated domains, and do not assume Z.ai's hosted API has the same data-retention terms as self-hosted weights.
"Does this change the US vs China model routing map?"
Yes, incrementally. A free stealth week proved demand; MIT weights plus $0.15/$0.50 API gives teams a tier-2 agentic route that does not depend on OpenRouter's anonymous provider. Pair with US frontier models for regulated tiers — same architecture AT&T documented.
Bottom line
GLM-5.3-Flash is Ox Alpha with a name, a license, and a price list. The mystery was fun; the product is what matters: 320B-A18B, 1M multimodal context, MIT weights, agent benchmarks competitive with Opus on GDPVal and DeepSWE, API at flash-tier pricing, and a claim of full stealth-week serving on Chinese silicon.
Run your evals. The stealth free lunch is over; the open-weight MoE era for agent harnesses is not.
Related on explainx.ai
- Ox Alpha forensics timeline — how the community got there first
- OpenRouter Ox Alpha setup guide
- Top 10 Ox Alpha use cases from stealth week
- GLM-5.3 launch — cyber-defense line (Aug 14)
- GLM-5.2 MIT open weights — prior cycle playbook
- China AI playbook — free models and cheap compute
- AT&T 56% coding cost savings from model routing
- Choosing open-weight vs closed models
- Hermes Agent — top Ox Alpha traffic sender
- Unsloth's 3-bit GLM-5.3-Flash GGUF — runs on 128GB RAM
- Qwen3.8-Flash goes live on QwenCloud — $0.16/$0.47 pricing, same Qwen4 preview
Official: Z.ai GLM-5.3-Flash blog · Hugging Face weights · @Zai_org launch post
Benchmarks, pricing, and architecture claims reflect Z.ai's August 26, 2026 launch materials. Run independent evals on your workloads before production routing. OpenRouter slug migration may lag the official announcement — verify live model IDs.