Anonymous Provider Launches Union Alpha, a Free 262K-Context Multimodal Model on OpenRouter
A third-party provider using the alias 'Stealth' has released Union Alpha on OpenRouter, a multimodal model with a 262K context window, currently free to use during its preview period. The model's developer remains anonymous, and OpenRouter states it is not the model's owner or operator.
Union Alpha Debuts as Free Preview on OpenRouter
A new multimodal model called Union Alpha appeared on OpenRouter on September 16, 2026, listed under the identifier stealth/union-alpha. The model's actual developer has chosen to remain anonymous during this preview period. OpenRouter, which routes inference requests to the model, has clarified it is not the model's developer, owner, or provider.
Union Alpha ships with a 262,000-token context window and is currently free to use through OpenRouter. According to the model's listing, it targets research, coding, and agentic workflows while aiming for what the provider calls "frontier-level performance across a broad range of general-purpose tasks." No independent benchmark scores, parameter count, or training cutoff date have been disclosed.
Performance Metrics
OpenRouter's own telemetry provides some operational data, though these figures reflect infrastructure performance rather than model quality:
- P50 latency: 5.39 seconds
- Throughput: 18 tokens per second
- Uptime (3-day): 99.85%
- Availability (3-day): 98.14%
These numbers describe how reliably OpenRouter can route and serve requests to the model, not its reasoning or coding capability. No standard benchmark results (MMLU, HumanEval, or similar) have been published for Union Alpha.
Data Handling
According to OpenRouter, prompts and completions sent to Union Alpha may be retained by the anonymous provider but are not used for training. All other usage is governed by OpenRouter's Stealth Model Terms, a separate policy applied specifically to unidentified model providers on the platform.
Part of a Broader Stealth Pattern
Union Alpha is not an isolated listing. The same anonymous provider (or providers operating under the "Stealth" branding) has also released Ox Alpha, a reasoning-focused model with a 1 million-token context window aimed at coding, long-horizon agentic work, and production use cases involving text and visual context. The pairing suggests a coordinated release strategy: one multimodal generalist model (Union Alpha) and one long-context reasoning specialist (Ox Alpha), both launched without public attribution.
Stealth releases on OpenRouter are a recognized pattern in the industry, often used by labs testing unreleased checkpoints under real-world traffic before an official launch and branding announcement. Free pricing during this phase is typical, functioning as an incentive for developers to generate usage data and feedback ahead of a formal release with paid tiers.
What This Means
Union Alpha's claims of "frontier-level performance" cannot be verified — there are no published benchmark scores, and the operational metrics available (latency, throughput, uptime) say nothing about output quality. The 262K context window is competitive but not exceptional against known models with 1M+ token windows. Until the provider identifies itself and publishes verifiable evaluation results, treat Union Alpha as an unproven entrant worth testing in low-stakes settings, not a model to build production systems around. The free pricing window is likely temporary and will end once the provider reveals its identity or moves to general availability.
Related Articles
Xiaomi open-sources Xiaomi-OCR-0, a 0.8B document model scoring 96.83 on OmniDocBench v1.6
Xiaomi has released the weights and inference code for Xiaomi-OCR-0, a 0.8B-parameter model that handles document parsing, key information extraction and OCR visual question answering. Xiaomi reports 96.83 on OmniDocBench v1.6, first among models under 1B parameters and second overall.
AgiBot launches GO-3 Preview and WITA-OMNI, topping three and four embodied AI leaderboards
AgiBot (智元) released two models at its Innovation Day: GO-3 Preview, which it says ranks first on three embodied intelligence benchmarks, and WITA-OMNI, which it says ranks first on four. It also introduced the Worken metric, the AHA3 hardware architecture, the G2 Max robot and Lingqu OS 1.0.
Microsoft releases Mage-VL-Cowork, a 4B computer-use agent scoring 66.0% on OSWorld-Verified
Microsoft has released Mage-VL-Cowork, a 4B-parameter vision-language policy that combines GUI control and terminal execution in one model. Microsoft claims it is the best 4B model on OSWorld-Verified, at 66.0% under a 50-step budget. Weights are published under Apache-2.0.
TBC launches Mischief, a Wan 2.2-based image-to-video model, on fal.ai October 30
The Biological Computing Co. (TBC) is launching TBC Mischief, an image-to-video model based on Wan 2.2, through fal.ai on October 30. TBC says it generates video five times faster at 80% lower cost than the underlying model, while improving output quality.
Comments
Loading...