Ox Alpha Was Free for a Week. The Weights Are MIT and 320B.

TL;DR

For roughly a week, an anonymous model called ox-alpha sat on OpenRouter and OpenCode with a million-token context, image input, and a price of zero. On August 26 Z.ai took the mask off: it was GLM-5.3-Flash, a 320B mixture-of-experts model with 18B active parameters, and the weights went up on Hugging Face under MIT the same day. Z.ai's published table puts it at 84.3 on Terminal-Bench 2.1 against Claude Opus 4.8's 85.0. List price is $0.15 in and $0.50 out per million tokens. Opus 4.8 on OpenRouter is $5.00 and $25.00.

The stealth week

Anonymous preview models are not new. What is new is one showing up free, with a 1M context and vision, and then turning out to be MIT-licensed rather than a closed API teaser.

Z.ai confirmed the connection first-party in its own launch material, which matters, because up to that point the identification rested on community tokenizer matching and leaked serving errors. The model card is unambiguous about what shipped: 320B total parameters, 18B active, and the first natively multimodal model in the GLM-5 series.

The usage receipts are visible on the OpenRouter model page. The top consuming application is Hermes Agent at 59.4 billion tokens, and the second is Claude Code at 40.3 billion. Anthropic's own CLI being the number two consumer of a Chinese open-weight model is either the sincerest compliment in the industry or a procurement conversation waiting to happen.

320 billion parameters, 18 billion awake

The efficiency story is the point of this release. GLM-5.3-Flash is the first GLM model to use a hybrid architecture that combines sparse and linear attention, which Z.ai says cuts long-context serving cost while keeping precise long-context behavior. It also adds Manifold-Constrained Hyper-Connections, a training-side change aimed at scaling efficiency, and it is pre-trained on a 30 trillion token multimodal corpus.

The reported effect on serving is a KV cache roughly 4.4 times smaller than GLM-5.3's full-attention design. That is the number that decides whether a 1M-token agent session is a product or a science experiment.

Worth being clear about what "18B active" does and does not buy you. It cuts the compute per token by about 18x versus a dense 320B. It does not cut the memory: you still have to house all 320 billion parameters somewhere, you just only wake 18 billion of them per token. It is renting the whole warehouse and lighting one aisle at a time.

The scores

Every number below is vendor-reported from Z.ai's own comparison table, so treat it as a starting hypothesis rather than a verdict.

The bigger jumps are against its own predecessor. On DeepSWE v1.1 it reports 63.4 against GLM-5.2's 46.2. On AutomationBench v1.0.6 it reports 48.8 against 26.2. On Humanity's Last Exam with tools it reports 55.3, roughly flat against GLM-5.2's 54.7 and behind Opus 4.8's 57.9.

Read that pattern honestly. Agentic coding and tool automation moved a lot. Hard reasoning barely moved. This is a model tuned for the loop you actually run, not the leaderboard you screenshot.

The price is the story

Z.ai frames it as one tenth the price of GLM-5.2. Against live OpenRouter list rates that is close enough: $0.15 and $0.50 versus $1.19 and $3.74. There is also a launch promotion running through September 9 at $0.075 in, $0.25 out, and $0.015 for cache reads, which is where the ratio against Opus stretches to about one hundredth.

The comparison that will actually change your bill is against the lab's own flagship. GLM-5.3 is $1.40 and $4.40. If your agent loop was running on GLM-5.3 because it was already the cheap option, the Flash variant is a roughly 9x cut on the same vendor, same API, same tokenizer.

It is not fast

"Flash" is doing some marketing work here. Artificial Analysis measures output speed at 50.2 tokens per second, against a median of 67.1 for open-weight models it tracks. It scores the model at 57 on its Intelligence Index, which puts it high among open-weight models of similar size, but the throughput sits below average.

So the name refers to cost per token and serving efficiency, not to how quickly a response streams back to a user. For batch agent work, evaluation runs, and long-context document jobs, that trade is fine. For anything with a human waiting on the first token, benchmark it yourself before you swap.

Read the footnotes

The most useful part of the model card is the section nobody quotes. Z.ai documents its evaluation harnesses, and the settings are generous:

• Terminal-Bench 2.1 was run inside Claude Code 2.1.207 with a six hour timeout and 65,536 max new tokens.
• DeepSWE was run through the mini-swe-agent harness at 400K context, temperature 0.95, and a six hour timeout.
• HLE with tools used a 300,000 token maximum context with a context management strategy, and GPT-5.6-luna at medium effort as the judge model.
• NL2Repo ran under 1M context with both rule-based and LLM-based judging specifically to block unauthorized pip and curl operations during the run.

None of that is misconduct. It is a lab publishing its setup, which is more than most do. But a six hour agentic timeout is not your CI budget, and a score produced inside a competitor's coding harness is a statement about the harness as much as the model. When you reproduce these, hold the timeout and the scaffold constant or you are measuring something else.

Running it yourself

The card lists first-party recipes for SGLang, vLLM, TokenSpeed, and KTransformers. That last one is the tell: KTransformers exists to run large MoE models with most experts parked in system RAM, which is the only realistic single-box path for 320B weights.

Do the arithmetic before you get excited. At 4-bit, 320B parameters is on the order of 160GB of weights before KV cache and activation overhead. The 18B active count makes it compute-cheap to run, not memory-cheap to host. MIT weights mean you can serve it commercially, fine-tune it, and distill from it without asking anyone. They do not mean it fits on your 5090.

What this actually changes

The playbook here is worth naming, because you will see it again. Ship the expensive flagship. Run the cheap sibling anonymously and free on the routers where agent developers live, collecting a week of real agentic traffic instead of synthetic evals. Then reveal it, open the weights, and price it under everything in its class.

For anyone building on top of models, the practical read is narrower. There is now an MIT-licensed model that a lab claims lands within a point of Opus 4.8 on terminal-agent work, at a token price two orders of magnitude lower, with the weights on disk so nobody can deprecate it out from under you. Whether it holds up in your harness is the only question left, and it is one you can answer this afternoon.

Key Takeaways

• The anonymous ox-alpha model on OpenRouter and OpenCode was GLM-5.3-Flash, confirmed first-party by Z.ai on August 26 and released under MIT the same day.
• 320B total parameters, 18B active, hybrid sparse plus linear attention, 30T-token multimodal pre-training corpus, and a KV cache about 4.4x smaller than GLM-5.3.
• Vendor-reported 84.3 on Terminal-Bench 2.1 against Opus 4.8's 85.0, with much larger gains over GLM-5.2 on DeepSWE (63.4 vs 46.2) and AutomationBench (48.8 vs 26.2).
• List price $0.15 in and $0.50 out per million tokens, roughly one fiftieth of Opus 4.8's output rate, with a promo through September 9 at $0.075 and $0.25.
• Artificial Analysis measures 50.2 output tokens per second, below the open-weight median of 67.1. "Flash" means cheap to serve, not fast to stream.
• Every benchmark figure is vendor-reported, and the published harnesses use six hour timeouts and 300K to 1M contexts. Reproduce with your own scaffold before you migrate anything.

Sources: Z.ai, GLM-5.3-Flash announcement, zai-org/GLM-5.3-Flash model card on Hugging Face, GLM-5 technical report (arXiv:2602.15763), OpenRouter GLM-5.3-Flash, Artificial Analysis, GLM-5.3-Flash, TestingCatalog, Hacker News discussion, Zapier AutomationBench