Hunter Alpha Was Xiaomi’s MiMo-V2-Pro—Its API Name Is Now Retired

Hunter Alpha’s identity is settled: the anonymous model was an early test build of Xiaomi’s MiMo-V2-Pro, not DeepSeek V4. The more consequential update is that MiMo-V2-Pro is no longer a current Xiaomi API model name; the company has replaced it with MiMo-V2.5-Pro.

That shift changes the practical meaning of the story. What began as speculation about an unidentified model now leads to a named, open-source successor whose weights and licensing terms are public, while the original V2-Pro endpoint has been retired.

How the DeepSeek theory ended

Hunter Alpha surfaced anonymously on a developer platform, presenting a capability profile that encouraged speculation about DeepSeek’s next model. A March 19 Reuters report identified it instead as an early internal test build of Xiaomi MiMo-V2-Pro and described the model as having one trillion parameters and a one-million-token context window.

The confusion joined three distinct propositions that should not be treated as equivalent. Hunter Alpha was the temporary public identity, MiMo-V2-Pro was the product Xiaomi connected to it, and DeepSeek V4 was an inference made before the developer became known. Similar output, architecture clues or personnel connections could support a hypothesis, but they could not establish ownership.

This distinction remains important because anonymous model testing reverses the usual launch sequence. Users encounter outputs, specifications and access conditions before they receive a model card, licensing terms or an accountable developer name. Early commentary can therefore become attached to a product identity that has not yet been verified.

What the headline specifications meant

The trillion-parameter figure conveyed scale, but not the amount of computation used for every token or a guaranteed level of performance. Mixture-of-experts systems can hold a much larger total parameter pool than the subset activated during an inference step, so total parameters and active parameters answer different technical questions.

Context capacity requires a similar qualification. A one-million-token limit means a model can accept an unusually large input, potentially accommodating extensive codebases, document collections or long agent sessions. It does not establish that every detail will be retrieved accurately, that reasoning quality remains constant across the full window or that a multistep task will avoid accumulating errors.

Nor did rapid activity on an aggregation platform function as a controlled evaluation. Usage can reflect promotional pricing, curiosity, routing decisions, repeated experimentation and useful performance in unknown proportions. It demonstrates demand for access, but not which tasks were completed successfully or how the model compared under identical tools and prompts.

MiMo-V2.5 became the relevant generation

Xiaomi moved the product family forward soon after the Hunter Alpha reveal. The company’s MiMo-V2.5 release notice records that public testing began on April 23, 2026, and that the weights were released under the MIT License on April 28, allowing commercial use, further training and fine-tuning without additional authorization.

The release consists of two principal models with one-million-token context support. MiMo-V2.5-Pro targets complex agent and coding work, while MiMo-V2.5 adds native understanding of text, images, video and audio. Those descriptions are product positioning from Xiaomi rather than independent proof that either model will perform reliably in every workflow.

The MIT-licensed weights nevertheless create a material difference from Hunter Alpha’s anonymous debut. Developers can inspect the published artifacts, retain a defined license and build deployments without depending exclusively on a temporary mystery listing. The ownership question is no longer the only available information about the model family.

Open weights do not remove the operational burden of a model at this scale. Commercial permission and technical feasibility are separate matters: inference hardware, memory, serving software, monitoring and tool integration can remain substantial costs even when the license permits modification and deployment.

MiMo-V2-Pro is now a historical endpoint

The status changed again in June. Xiaomi’s official deprecation table lists June 30, 2026, as the retirement date for the mimo-v2-pro API name and identifies mimo-v2.5-pro as its replacement; requests had begun routing to the successor on June 1.

MiMo-V2-Pro should therefore be described as the model behind Hunter Alpha, not as Xiaomi’s current flagship endpoint. The identity remains historically significant, but anyone assessing present availability needs to distinguish the retired API name from the later weights and service options.

This also prevents a misleading interpretation of the open-source update. Xiaomi did not retroactively turn the anonymous Hunter Alpha listing into a permanent public product. It introduced a subsequent generation with named variants, published licensing terms and a migration path away from the V2 model names.

Why the episode still matters to creators

For creators evaluating AI systems, the episode illustrates why model identity, platform metadata and demonstrated performance need separate treatment. A platform listing can establish what users were offered at a particular moment, while a developer publication can define architecture, intended workloads and licensing. Neither substitutes for an independent evaluation conducted with disclosed prompts, tools and scoring conditions.

The same separation applies to practical model selection. Long context may be relevant to transcript archives, source packets and multi-file production work, while agent optimization may benefit processes involving repeated tool calls. Those features do not answer whether a model cites sources correctly, follows editorial constraints, preserves confidential material or produces consistently usable copy.

Parameter count is equally unsuitable as a shortcut for those judgments. It describes part of a system’s architecture, not its factual reliability, editorial control or total deployment cost. The appropriate comparison depends on the complete configuration, including the model version, agent framework, tools, retrieval setup and human review.

Hunter Alpha’s mystery has consequently become secondary to the model family’s documented lifecycle. It was Xiaomi’s MiMo-V2-Pro test identity; MiMo-V2-Pro was later displaced at the API level; and MiMo-V2.5 now provides the named, MIT-licensed branch that developers can evaluate directly.

Also read:

• How’s Allbirds Doing After Its Wild AI Pivot? Now It’s Smartbird — and the Sneakers Are Officially History
• Substack Now Lets Readers Check If a Newsletter Was Written by AI
• Blog Link Building Still Works—But Its Old Shortcuts Now Count as Spam