Stealth Model Ox Alpha Stuns Community; Fingerprints Point to Zhipu GLM-5.3

An anonymous model called Ox Alpha suddenly appeared on OpenRouter and OpenCode this week with no vendor attribution, no paper, and no announcement, yet it quickly blew up in the community with a 1M token context window, multimodal text/image/video input, a zero-data-retention policy, and a near-unlimited free quota for a week—its true identity and origin remain the biggest mystery.

Confirmed

• Strong real-world performance: Ox Alpha scored around 80% across 10 DeepSWE tasks, beating GPT-5.6-sol (52%) and Fable (65%); Teknium tested OX Alpha Max in Hermes Bot and called its performance excellent and unusually intelligent, with front-end quality beyond expectations.
• @teortaxesTex's review found the model surpasses Kimi K3, extremely fast with a Claude-like experience, and half-jokingly suspected it was a stolen Claude checkpoint.
• @FlunkyGraphics ran three black-box fingerprint tests (including tokenizer consistency comparisons across 6 languages) and confirmed Ox Alpha is based on the GLM-5.3 architecture.
• A system prompt leaked by @elderplinius shows the model is forced to claim it is "ox-alpha," developed by an undisclosed organization, when asked about its identity—strictly forbidden from revealing its true identity—and points to Zhipu.
• Hosting platform OpenCode claims compute capable of processing 100 trillion tokens per day.
• Known weakness: @bindureddy noted that while the model is free, long-context, and multimodal, it tends to produce idle filler content, so real-world usability is questionable.

Unconfirmed

• The model's ultimate ownership has not been officially confirmed. The mainstream guess is Zhipu's GLM family, possibly a GLM 5.3 vision variant or GLM 5.3 Flash; a message relayed by @calabiandyau also claims Kimi k3.1 is coming soon.
• The model once claimed during testing to be Mistral's "Le Chaton Fat," contradicting the GLM hypothesis—clearly a result of system-prompt constraints or unreliable self-reporting.
• @teortaxesTex challenged the "Flash small model" claim: its speed is slow for a Flash-tier model, though this could relate to high batch-size inference; others think it's more like a 5.x release with enhanced vision.
• Notable blogger Elvis (@omarsar0) said insiders have confirmed the true identity (which he can't disclose yet) and called it "spectacular"-tier, but no details were shared.

Why it matters

• If the GLM-5.3 fingerprint conclusion holds, this would be a landmark case of a Chinese frontier model stealthily benchmarking itself directly against top models like GPT-5.6 and Fable; @teortaxesTex also pre-registered the judgment that it's very likely a Zhipu product and argued the multimodal agent race timeline needs to be moved up overall.
• The leaked "hide your identity" system prompt has sparked discussion about stealth evaluations, marketing tactics, and model trustworthiness.
• @Acceptable-Object390 used the open-source agent framework Row-Bot to complete a self-built self-check pipeline with 20 tool calls from a single prompt, demonstrating its practicality on long-horizon agent tasks.

2026-08-21 ~ 2026-08-23 · 55 related posts

• OpenCode Offers Free Access to Ox Alpha: 1M Context and Multimodal — TJLarkin23 · 2026-08-21
• Ox Alpha Review: Faster than Kimi K3, Rivals Claude — teortaxesTex · 2026-08-21
• OpenRouter's stealth model Ox Alpha sparks guesses about Chinese AI vendors — lxfater · 2026-08-21
• Mystery Model Ox Alpha Suspected to be Zhipu GLM — teortaxesTex · 2026-08-21
• New Ox Alpha model leaked with 1M context and multimodal input — External_Mood4719 · 2026-08-21
• Rumor: Ox Alpha is a joint ZAI and GLM model — scaling01 · 2026-08-21
• Stealth Model 'Ox Alpha' Hits 80% on DeepSWE Benchmark — AccBalanced · 2026-08-21
• Speculation suggests OpenAI o1 uses GLM architecture, rivaling Claude — teortaxesTex · 2026-08-21
• [source] Mystery Model OX Alpha Max Impresses with Unlimited Access via OpenRouter — Teknium · 2026-08-21
• Tests suggest O1 Alpha is a GLM-based model with advanced coding capabilities — teortaxesTex · 2026-08-21
• Lab praised for high compute efficiency and rapid model delivery — teortaxesTex · 2026-08-21
• Stealth Model Ox-Alpha Released, Outperforms Fable on SWE Benchmark — troll_khan · 2026-08-21
• Mysterious Lab Expected to Rapidly Deliver Frontier-Level SWE Models — teortaxesTex · 2026-08-21
• Internal Models Strong Yet Rely on Older US Base Models — teortaxesTex · 2026-08-21
• GLM 5.3 "Flash" spotted internally; observers doubt it's a small model — teortaxesTex · 2026-08-21
• Mystery stealth model 'Ox Alph' appears on OpenRouter, China lab suspected — Neosinic · 2026-08-21
• Ox Alpha open-source model benchmarks near Anthropic level — liminal_bardo · 2026-08-21
• Ox Alpha Stealth Model Leaked with 1M+ Context & Multimodal Capabilities — WorldofAI · 2026-08-21
• Users Praise Mystery Model '0xalpha' as Absurdly Capable — cephaloform · 2026-08-21
• Speculation: ox-alpha Model Might Actually Be GLM 5.3-vision — teortaxesTex · 2026-08-21
• Community bets mystery model Ox Alpha is Zhipu's GLM 5.3 vision variant — teortaxesTex · 2026-08-21
• Ox Alpha on OpenRouter identified as GLM-5.3 variant with alignment tax — airesearch12 · 2026-08-21
• Multimodal model ox-alpha launches on AGIBar with unlimited 1-week testing — 赛博禅心 · 2026-08-21
• User praises Ox Alpha: Fast and better than Kimi K3 in some ways — jd_pressman · 2026-08-21
• Anonymous "Ox Alpha" model researches, builds and QA's its own site in one prompt — Acceptable-Object390 · 2026-08-21
• [source] Black-box fingerprint confirms Ox Alpha is based on GLM-5.3 — FlunkyGraphics · 2026-08-21
• Ox-alpha solves coding task others failed; GLM-5V matches token count — eyishazyer · 2026-08-21
• Mystery model ox-alpha likely Zhipu's GLM: token fingerprints match GLM-5.3 exactly — eyishazyer · 2026-08-21
• Zhipu's GLM-6 Leaks: reportedly Outperforms Frontier Models in SWE Benchmarks — hunarbatra · 2026-08-21
• Mystery Model OX ALPHA Appears on OpenRouter with Top-Tier Coding Benchmarks and No Filters — heypearlai · 2026-08-21
• OpenRouter's stealth model Ox Alpha sparks guessing game over its identity — AccBalanced · 2026-08-21
• Mystery Model 'Ox Alpha' Appears on OpenRouter with 1M Context — basedjensen · 2026-08-21
• [source] Insider confirms stealth model Ox Alpha is spectacular, resets AGI timelines — omarsar0 · 2026-08-21
• Ox Alpha is a Chinese model — Bindu Reddy bets it's Xiaomi, not GLM — bindureddy · 2026-08-21
• Ox Alpha: Free Chinese model with 1M context but high latency — bindureddy · 2026-08-21
• Rumor: Zhipu's GLM 5.3 Flash and Kimi k3.1 incoming on OpenRouter — calabi_and_yau · 2026-08-21
• Rumor: Ox Alpha might be based on Zhipu GLM 5.3 Flash — andrew_n_carr · 2026-08-22
• Mystery Model Ox Alpha Review: Near-Fable Performance, Unique Vision/Logic Profile — Afinetheorem · 2026-08-22
• Mystery model 'Ox alpha' appears on OpenRouter and OpenCode — altryne · 2026-08-22
• Hands-on with Ox Alpha: Impressive Performance in Pi Harness — omarsar0 · 2026-08-22
• Mystery Model ox-alpha Leaks: Open Reasoning Traces, No Vision Support — zainhas · 2026-08-22
• Mystery Model OxAlpha Reveals Itself as Mistral's Le Chaton Fat — rickasaurus · 2026-08-22
• Mysterious 'Ox Alpha' model offers 1M context, huge compute claims — udmrzn · 2026-08-22
• Leaked System Prompt Shows GLM-5.X 'Ox-alpha' Hiding Its True Identity — Yamapama · 2026-08-22
• Ox Alpha free with high limits on OpenRouter, ideal for batch inference tasks — mhmazur · 2026-08-22
• Stealth Model 'Ox Alpha' Launched with 1M Context and Free Access — 1littlecoder · 2026-08-22
• GLM 5.3 compression plus vision rivals oX performance — teortaxesTex · 2026-08-22
• Mysterious OxAlpha model debuts with 1M context and free access — APPSO · 2026-08-22

2 near-duplicate retellings: heypearlai · bindureddy