AI Wiki

Explore AI Wiki

Citation and evidence

Ox Alpha

15 min full readUpdated 24 references

This article's verification

Report a problem with this article

More

Use this article

Raw MarkdownExplore connections

Improve this page

Suggest editRevision historyDiscussion

Browse categories

AI Code GenerationAI ModelsLarge Language Models

Cite this article

Ox Alpha was the anonymous preview alias for GLM-5.3-Flash, a natively multimodal large language model developed by Z.ai. Z.ai tested the model anonymously through OpenRouter and OpenCode, where it became available on Aug. 20, 2026. OpenRouter originally described it as a reasoning model for coding and sustained agent work without naming the provider.[1][2][21]

The public preview listing documented a 1,048,576-token context window, text, image, and video input, text output, and a maximum output of 131,072 tokens on the OpenRouter route. OpenCode announced a one-week free preview with a separate route and privacy policy.[3][5] Those were distribution and interface facts. They did not, by themselves, identify the developer or disclose the architecture, parameter count, training process, weights, or license. Z.ai's Aug. 26 release and model repository later supplied the developer identity, architecture summary, parameter counts, public weights, and MIT license. The release post and model card do not include a full training-corpus inventory or training-compute record.[21][22]

Through Aug. 22, neither distributor had named the developer or linked a developer-authenticated checkpoint, technical report, model card, or reproducible evaluation package. Ox Alpha therefore remained a temporary stealth preview at that time.[2][6] Z.ai resolved the identity on Aug. 26 when it released GLM-5.3-Flash and said that ox-alpha had been its anonymous test alias.[21] GLM-5.3-Flash is distinct from GLM-5.3, Z.ai's larger text-only flagship announced earlier in August.[21][24]

Key facts

FieldDetail
Public previewAug. 20, 2026[1][5]
Revealed identityGLM-5.3-Flash, announced Aug. 26, 2026[2][21]
DeveloperZ.ai; undisclosed during the preview[2][21]
OpenRouter model IDstealth/ox-alpha[3]
OpenCode Zen model IDx-preview-f-free[6][7]
OpenCode Go model IDox-alpha-free[8]
Stated focusCoding, sustained agentic work, and workflows combining text with visual context[1][2]
Context window1,048,576 tokens on the listed routes[3][5]
Maximum output131,072 tokens on OpenRouter[3][4]
Input and outputText, image, and video input; text output[3]
OpenRouter reasoning controlMandatory; max, high, and low effort settings[3]
OpenRouter and Zen token priceFree at the research cutoff[2][6]
OpenCode Go accessGo subscription ($5 for the first month, then $10 per month); Ox Alpha is listed without token charges for a limited time[8]
Public artifact statusPublic weights, model card, and checkpoint configuration under the MIT license; no Flash-specific technical report at the research cutoff[21][22]

Expanded article table

Preview and distribution

Identity reveal

Z.ai's Aug. 26 release post says the company tested GLM-5.3-Flash anonymously as ox-alpha on OpenCode and OpenRouter to gather user feedback. OpenRouter then updated the Ox Alpha page with a notice identifying ZAI as the developer and operator and naming GLM-5.3-Flash as the revealed model.[2][21] This explicit first-party statement and distributor relabeling supplied the evidence that had been absent during the anonymous period. It did not validate preview-era guesses about the developer or make the model identical to the separate GLM-5.3 flagship.

OpenRouter route

OpenRouter's announcement presented Ox Alpha as a stealth model for AI code generation and long-running agent workflows. The launch post gave two main specifications: a one-million-token context window and text, image, and video input. A follow-up in the same thread said the route was free and that the provider did not train on prompts or completions.[1]

During the anonymous phase, the model page said a third-party provider developed and operated Ox Alpha and that OpenRouter only routed requests. The updated page preserves that preview description but adds a notice identifying ZAI as the developer and operator and linking the revealed GLM-5.3-Flash model.[2] The endpoints API used the generic provider name Stealth during the preview; that route label did not disclose the company at the time.[4]

OpenRouter priced both prompt and completion tokens at zero during the preview. Its stealth terms make clear that this category of access is time-limited and may be removed with or without notice.[2][9] The zero price was therefore a dated preview condition, not the released model's permanent price or a promise that the route would continue.

OpenCode routes

OpenCode announced that Ox Alpha would be free for the next week. Its launch post repeated the one-million-token and multimodal-input claims and advertised generous rate limits, near-unlimited use, zero data retention, and capacity for 100 trillion tokens per day.[5]

That final figure is an operator capacity claim. It is not a count of tokens actually processed, a per-user allocation, a measured tokens-per-second rate, a service-level guarantee, or evidence about the model's parameter count. OpenRouter publishes changing traffic and performance counters for the route, but those short-window measurements are different from OpenCode's aggregate capacity statement and are not stable model specifications.[2][5]

OpenCode Zen listed Ox Alpha's input, output, and cached-read prices as free. OpenCode Go has a separate plan charge: $5 for the first month and $10 per month afterward. The Go catalog showed no token charges for Ox Alpha and labeled it free for a limited time.[6][8] The word "free" therefore described the model's preview token price, not the absence of every possible subscription charge.

OpenCode uses different identifiers in its two product catalogs. Its Zen documentation maps Ox Alpha Free to x-preview-f-free on an OpenAI-compatible chat-completions endpoint.[6][7] OpenCode Go maps the same preview name to ox-alpha-free on its own endpoint.[8] The owned_by: opencode value in OpenCode's model API identifies the API namespace. It does not say that OpenCode trained the underlying model.[7]

Launch traffic on OpenRouter

OpenRouter compared Ox Alpha with other model launches using token volume in each launch's first three full UTC days. Its Aug. 24 chart reported 11.6 trillion tokens for Ox Alpha over Aug. 21 through Aug. 23 and described that rounded total as 2.6 times the next-highest launch in the comparison, Qwen3.6 Plus at 4.4 trillion tokens.[17] OpenRouter's public model activity data sum to 11,556,229,871,598 prompt and completion tokens for those three completed days, which explains the rounded chart value.[2]

The same post said Ox Alpha was on track to process nearly 6 trillion tokens on Aug. 24; that statement was a forecast, not a completed-day result.[17] OpenRouter's later completed UTC record lists 5,835,184,092,873 prompt tokens and 94,402,126,602 completion tokens, or 5,929,586,219,475 total, across 70,334,611 API requests for that day.[18]

The detailed activity record marked 5,348,751,464,576 of the Aug. 24 prompt tokens as cached reads. OpenRouter defines cached tokens as tokens read from cache and says they are included within prompt tokens rather than added to them.[2][19] The request count is not a count of unique users. These figures describe traffic through OpenRouter, use provider-native token counts, and do not establish model quality or usage through other routes.[18][20]

Documented interface and limits

The OpenRouter models API describes a text, image, and video to text route. It lists a 1,048,576-token context length and a 131,072-token maximum completion. The endpoint accepts controls for maximum tokens, sampling temperature, top-p and top-k sampling, response format, tools, tool choice, and reasoning effort.[3][4]

Reasoning is mandatory on this route, enabled by default, and exposed at max, high, or low effort. The listed defaults are temperature 1 and top-p 0.95.[3] These fields describe how a client can call the hosted route. During the preview, they did not show whether the underlying model was dense or sparse, how it was trained, how many parameters it had, or how its internal reasoning process worked. The later release and checkpoint documented a sparse mixture-of-experts architecture and the model's parameter counts, but not its internal reasoning process.[21][22]

Tool and tool-choice parameters make the route usable in AI agents that implement function calling. Parameter support alone does not establish reliable tool selection, correct structured output, or success on long software-engineering tasks. Those behaviors require controlled evaluation with a specified agent, environment, and grader.[3][4]

The modality field has a similarly narrow meaning. The hosted route fits the interface definition of a multimodal model: Ox Alpha accepts text, images, and video and returns text. The listing does not document audio input, audio output, image generation, or video generation.[3] It also does not state how a video is sampled, whether its audio track is processed, or how visual tokens count against the context budget.

Data handling by route

Ox Alpha's access routes have different published data-handling statements. They should not be collapsed into one model-wide privacy claim.[2][5][6][8]

RouteRetention statementTraining statementScope
OpenRouterThe provider, anonymous at launch and later identified as Z.ai, retains prompts and completionsThe route page says they are not used; the incorporated EULA authorizes training and improvementOpenRouter route and Stealth Program terms[2][9][21]
OpenCode ZenZero retentionNot used for model trainingOpenCode route[5][6]
OpenCode Go0 daysNot used for model trainingOpenCode Go route[8]

Expanded article table

OpenRouter's route-specific notice does not amount to zero data retention, and the route page and incorporated EULA do not align cleanly on training. The Ox Alpha page says the provider retains prompts and completions but does not use them for training.[2] The EULA says Stealth Models collect user content for training and improvement, tells users who do not want that use to refrain from accessing them, and grants OpenRouter and the provider rights to use content for training, evaluation, and improvement. It also says user content may be collected and shared with the provider and that personal data in an input will be sent to it.[9] The published materials do not explain how the Ox Alpha-specific notice limits those broader terms. The supportable conclusion is nonzero retention and an unresolved training-policy conflict, not an unqualified no-training guarantee.

OpenCode's zero-retention statement applies to access through OpenCode. It cannot be used to describe an OpenRouter request, even when both routes reach the same preview name. During the anonymous period, users also could not compare the unnamed provider's internal controls, audit reports, or jurisdiction with those of a named model vendor.[2][6][8] The Aug. 26 reveal resolved the developer's identity, but it did not retroactively align the two routes' data policies or publish the missing controls and audits.[21]

Evaluation evidence

No official Ox Alpha benchmark report accompanied the preview. Ben Davis posted an informal run of ten DeepSWE tasks and explicitly warned that such a small subset could have substantial variance.[10] The post did not provide the task IDs, random seed, benchmark version, agent configuration, model settings, complete trajectories, grader files, or full logs. Its resulting percentage cannot be treated as a model-wide DeepSWE score or a controlled head-to-head result.

DeepSWE contains 113 original, long-horizon software-engineering tasks across 91 active open-source repositories and five programming languages. Each task has a hand-written verifier intended to check requested behavior rather than one reference patch.[11] The public repository uses the Harbor task format and a separate verifier environment. It documents deterministic subset sampling and says its leaderboard runs use Pier with mini-swe-agent on Modal, with evaluation trajectories released for inspection.[12]

Those details matter because a coding-agent result measures a full system, not only model weights. The agent scaffold, tools, network access, time limit, token budget, sampling settings, retries, selected tasks, benchmark revision, and verifier execution can all affect the outcome. Comparisons produced under different configurations are not interchangeable. DeepSWE's relationship to SWE-bench does not remove that requirement; it was designed in part to improve task originality and grading, not to make undocumented samples representative.[11]

At the preview cutoff, no full Ox Alpha DeepSWE trajectory package, official 113-task submission, or independent replication was located. The preview could be tested by users, but anecdotes and screenshots did not establish a benchmark rank.[10][12] Z.ai later published an official benchmark suite under the GLM-5.3-Flash name. Those vendor results are documented with the released model and do not turn the earlier ten-task sample into a controlled full-benchmark result.[21]

Interpreting context and multimodal claims

A context-window number states how much tokenized input a route accepts. It does not show how reliably the model finds or combines information throughout that input. "Lost in the Middle" found that long-context language models often performed best when relevant information appeared near the beginning or end and worse when it appeared in the middle.[13] RULER extended long-context testing beyond simple retrieval to multi-hop tracing and aggregation, and reported substantial performance declines as context length and task complexity increased.[14]

Neither study tested Ox Alpha. They establish why the one-million-token listing should be described as capacity rather than proof of effective reasoning over one million tokens. A direct evaluation would need controlled tasks at multiple lengths and positions, with the exact route and settings recorded.[3][13][14]

Video input also needs task-level evaluation. LongVideoBench tests retrieval and reasoning over interleaved video and language inputs up to one hour, while Video-MME covers short, medium, and long videos across several content domains.[15][16] No Ox Alpha result on either benchmark was located. The supported-input field therefore establishes that a client may send video, not that the model understands long temporal sequences at a particular level.

Identity reveal and remaining disclosure limits

During the preview, OpenRouter labeled the API tokenizer as Other, left the knowledge cutoff unset, provided no Hugging Face repository ID, and labeled quantization as unknown.[3][4] None of those route fields identified a model family. Similar API parameters or tokenization behavior can occur across related serving stacks and did not establish authorship. Z.ai's Aug. 26 statement and OpenRouter's explicit relabeling provided the missing identity evidence.[2][21]

The release resolved several disclosure gaps while leaving others open.[21][22][23]

AreaStatus after the release
Developer identityZ.ai, which says it ran the anonymous preview[2][21]
ArchitectureZ.ai identifies a sparse mixture-of-experts model with 320 billion total parameters, 18 billion active parameters, and hybrid linear and sparse attention[21][22]
TrainingZ.ai reports a newly trained base and a 30-trillion-token multimodal pre-training corpus, but it does not publish the corpus inventory, full recipe, training compute, or knowledge cutoff[21][22]
ArtifactsPublic weights, model card, and checkpoint configuration under the MIT license; no Flash-specific technical report was published at the cutoff[22]
EvaluationZ.ai published a benchmark suite, but complete raw outputs, a public safety report, and independent replication were not supplied for every result[21][22]
OperationsZ.ai says preview inference ran on Chinese AI chips, but it does not name the accelerator model or publish an independent serving audit[21]

Expanded article table

Community attempts to infer the developer from output wording, model self-identification, error strings, geopolitical answers, and tokenizer comparisons were not sufficient evidence during the anonymous phase. The first-party announcement and OpenRouter relabeling now establish the identity. They do not retroactively validate any specific preview-era guess.[2][3][4][21]

Availability and reproducibility

OpenCode framed its initial offer as one week. OpenRouter's stealth terms say availability is limited and may end at the provider's request or OpenRouter's discretion.[5][9] Those statements describe the initial preview route rather than the durability or terms of the named model release.

Z.ai has now published weights, a model card, and checkpoint configuration under the MIT license.[21][22] Researchers can inspect and self-host the released checkpoint, but they cannot reproduce its training from the public record because the full dataset, recipe, and compute record remain undisclosed. Historical black-box results also remain dependent on the mutable preview endpoints. The identity reveal supports linking Ox Alpha to GLM-5.3-Flash; it does not establish undocumented route behavior or validate earlier speculation.[2][3][6][9][21][22][23]

References

  1. ^1 ^2 ^3 ^4OpenRouter. "New stealth model: Ox Alpha." X, Aug. 20, 2026. x.com/...2090544970923184269
  2. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12 ^13 ^14 ^15 ^16 ^17 ^18 ^19 ^20OpenRouter. "Ox Alpha: API Pricing & Providers." Accessed Aug. 27, 2026. openrouter.ai/...ox-alpha
  3. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12 ^13 ^14OpenRouter. "Models API." Ox Alpha entry, accessed Aug. 22, 2026. openrouter.ai/...models
  4. ^1 ^2 ^3 ^4 ^5 ^6OpenRouter. "Ox Alpha Endpoints API." Accessed Aug. 22, 2026. openrouter.ai/...endpoints
  5. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8OpenCode. "Ox Alpha (stealth model) is free for the next week." X, Aug. 20, 2026. x.com/...2090544355824038300
  6. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9OpenCode. "Zen." Accessed Aug. 22, 2026. opencode.ai/...zen
  7. ^1 ^2 ^3OpenCode. "Zen Models API." Accessed Aug. 22, 2026. opencode.ai/...models
  8. ^1 ^2 ^3 ^4 ^5 ^6 ^7OpenCode. "Go." Accessed Aug. 22, 2026. opencode.ai/...go
  9. ^1 ^2 ^3 ^4 ^5OpenRouter. "Stealth Program End User License Agreement." Updated July 6, 2026. openrouter.ai/...stealth
  10. ^1 ^2Ben Davis. "I ran this thing through 10 tasks on DeepSWE." X, Aug. 21, 2026. x.com/...2090655207831298095
  11. ^1 ^2Wenqi Huang, Charley Lee, Leonard Tng, and Serena Ge. "DeepSWE: Measuring Frontier Coding Agents on Original, Long-Horizon Engineering Tasks." arXiv:2607.07946, July 8, 2026. arxiv.org/...2607.07946
  12. ^1 ^2Datacurve AI. "DeepSWE." GitHub repository, accessed Aug. 22, 2026. github.com/...deep-swe
  13. ^1 ^2Nelson F. Liu et al. "Lost in the Middle: How Language Models Use Long Contexts." Transactions of the Association for Computational Linguistics, 2024. arxiv.org/...2307.03172
  14. ^1 ^2Cheng-Ping Hsieh et al. "RULER: What's the Real Context Size of Your Long-Context Language Models?" COLM 2024. arxiv.org/...2404.06654
  15. ^Haoning Wu, Dongxu Li, Bei Chen, and Junnan Li. "LongVideoBench: A Benchmark for Long-context Interleaved Video-Language Understanding." arXiv:2407.15754, July 22, 2024. arxiv.org/...2407.15754
  16. ^Chaoyou Fu et al. "Video-MME: The First-Ever Comprehensive Evaluation Benchmark of Multi-modal LLMs in Video Analysis." arXiv:2405.21075, revised May 30, 2025. arxiv.org/...2405.21075
  17. ^1 ^2OpenRouter. "Ox Alpha is on track to hit nearly 6 trillion tokens today." X, Aug. 24, 2026. x.com/...2091912024922177562
  18. ^1 ^2OpenRouter. "Model rankings daily data." Public rankings API, Aug. 24, 2026 record, accessed Aug. 25, 2026. openrouter.ai/...models
  19. ^OpenRouter. "Usage Accounting." Accessed Aug. 25, 2026. openrouter.ai/...usage-accounting
  20. ^OpenRouter. "Daily Token Totals for Top 50 Models." Accessed Aug. 25, 2026. openrouter.ai/...ly-token-totals-for-top-50-models
  21. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12 ^13 ^14 ^15 ^16 ^17 ^18 ^19 ^20 ^21 ^22Z.ai. "GLM-5.3-Flash: Frontier Intelligence, Flash Cost." Aug. 26, 2026. z.ai/...glm-5.3-flash
  22. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10Z.ai. "zai-org/GLM-5.3-Flash." Hugging Face model card and weights, accessed Aug. 27, 2026. huggingface.co/...GLM-5.3-Flash
  23. ^1 ^2Z.ai Developer Documentation. "GLM-5.3-Flash." Accessed Aug. 27, 2026. docs.z.ai/...glm-5.3-flash
  24. ^Z.ai. "GLM-5.3: Frontier Coding with Emergent Cyber Capabilities." Aug. 14, 2026. z.ai/...glm-5.3

Improve this article

Add missing citations, update stale details, or suggest a clearer explanation. Every suggestion is reviewed for sourcing before it goes live.

2 revisions · v3 · 2,939 words · full history

Fact-checks are independent of edits: a reviewer re-verifies the article against its sources and stamps the date. How we verify

Research and drafting on this wiki are AI-assisted, under named human editorial standards. How AI is used here

Reviewer note: Independently checked against primary, technical, academic, and corroborating sources through 2026-08-22.

Cite this page: AI Wiki. "Ox Alpha." aiwiki.ai, updated 27 Aug 2026, fact-checked 22 Aug 2026. CC BY 4.0. https://aiwiki.ai/wiki/ox_alpha

Suggest edit

What links here

GLM-5.3-Flash