Uncensored MiMo Vision API

Cheap multimodal endpoints are cheap because something was cut, and what gets cut is usually the part you cannot see: the weights are squeezed to three or four bits, and the model still answers confidently about a chart it can no longer actually read. MiMo-V2.5 is the counter-example in our catalog — at the price floor and at the top precision tier simultaneously, because the foundation it descends from was built small and dense rather than enormous and sparse.

01Where MiMo comes from: a reasoning model first, a vision model second

The MiMo line is Xiaomi's open-weights research program, and its founding argument is unfashionable: reasoning ability is set during pretraining, not bolted on afterward with RL. The work is mostly data engineering — Xiaomi rebuilt text extraction and filtering to raise the density of reasoning patterns in the corpus, generated synthetic reasoning data at scale, and pretrained MiMo-7B-Base on roughly 25 trillion tokens under a three-stage mixture, with Multiple-Token Prediction as an auxiliary objective that doubles at inference time as a speculative decoder (acceptance rate near 90%, per Xiaomi).

Posttraining is deliberately austere: 130,000 math and code problems with rule-based accuracy rewards only, no learned reward model, on the reasoning that a neural reward model is a surface to be hacked. Verifiable rewards cannot be sweet-talked. MiMo-7B-RL posts numbers out of proportion to its size — Xiaomi reports 68.2% on AIME 2024, 95.8% on MATH-500, 57.8% on LiveCodeBench v5 at pass@1, matching OpenAI's o1-mini. That is the shape of the model you are calling: compact, densely trained, tuned to work through a problem rather than recall an answer — and that temperament survives into the multimodal descendants.

02How MiMo learned to see: the MiMo-VL architecture

MiMo-VL is the vision-language branch, and its construction is instructive. Xiaomi assembled three parts: a native-resolution ViT encoder that preserves fine detail instead of forcing every image through a fixed low-resolution grid, a randomly initialized MLP projector into the language model's latent space, and — critically — the MiMo reasoning backbone as the language model. The vision system inherits a decoder already good at multi-step thinking rather than one merely good at captioning.

Training ran in four pretraining stages — projector warmup, vision-language alignment, general multimodal pretraining, long-context SFT — across roughly 2.4 trillion tokens, then Mixed On-policy Reinforcement Learning (MORL), folding perception accuracy, visual grounding precision, reasoning correctness, and human/AI preference into one on-policy loop. The report is candid that optimizing all of those at once is unstable — a more honest framing than most multimodal papers offer.

The published results are strong for the parameter count. Xiaomi reports MiMo-VL-7B-RL beating Qwen2.5-VL-7B on 35 of 40 tasks, 59.4 on OlympiadBench (above models up to 78B), and 56.1 on OSWorld-G for GUI grounding, ahead of purpose-built interface agents. The later 2508 refresh pushed MMMU past 70 for the first time in the series (70.6), reached 94.4 on ChartQA, 92.5 on ScreenSpot-v2 and 70.8 on VideoMME, and added a thinking-control switch so reasoning traces can be turned off per request.

03What W8A16 means, and why it matters more for vision than for text

Every model on the gateway carries a quantization suffix. It is not decoration — it tells you what was traded to make the model servable.

Suffix	What it is	Why it gets used
3BIT-REAP	3-bit weights plus REAP, an expert-pruning method that drops rarely-routed experts from an MoE model before quantizing	The only way to fit and serve the very largest foundations at a sane speed
W4A16-AUTOROUND-REAP	4-bit weights, 16-bit activations, quantized with AutoRound — learned rounding that recovers accuracy over naive round-to-nearest — plus REAP pruning	Middle ground for large foundations that do not need extreme compression
W8A16	8-bit weights, 16-bit activations — the highest-fidelity tier we serve	Used where the foundation is small enough that aggressive compression is unnecessary

MiMo-V2.5 earns the third row by being compact: a dense foundation in this class does not need crushing to three bits to fit, so you can afford to keep eight, and the weights you sample from stay close to the ones that were trained. That matters disproportionately for image work, for a structural reason. In a text-only decoder, an embedding that drifts under quantization gets re-contextualized by dozens of later attention layers — many chances to recover. A vision stack has no such slack: visual information passes through a single encoder and a single narrow projector before it reaches the decoder, and everything downstream is conditioned on what that bottleneck produced. Fine-grained signal — 3 versus 8 in a low-contrast scan, the exact pixel where a bar ends on a chart axis, the boundary of a UI element you are asked to click — lives in the low-magnitude weight detail that aggressive quantization discards first.

A 3-bit vision model does not fail loudly; it fails quietly, producing a fluent description of a document with two digits transposed in the total. Higher weight precision is the cheapest insurance against that failure, and here it is free — you pay the floor price either way.

04The floor price, in context

Four models in our twelve-model catalog sit at the price floor of $0.50 input / $2.00 output per million tokens. Only one of them takes images. Here is where the vision-capable models land on output cost, the number that dominates most real bills:

Output price in USD per 1M tokens, Shannon AI Gateway list rates, September 2026. Bar length is relative to the most expensive vision model in the catalog.

Against Kimi-K3 — the ceiling vision model, and a genuinely stronger one — MiMo-V2.5 is roughly 9.6× cheaper on output and 7.7× cheaper on input. That is not a claim of equivalence; it is an argument about workload shape. At a hundred thousand images a month, the ceiling model is not affordable and the floor model is, and the honest question is whether the floor model clears your accuracy bar — not whether it wins a benchmark.

05MiMo-V2.5 vs MiniMax-M3: the two cheap vision models, compared honestly

These two tie exactly on price, context, and endpoint surface. They are not interchangeable: the quantization tier is a proxy for two different bets about what makes a vision model useful.

	MiMo-V2.5 · W8A16	MiniMax-M3 · 3BIT-REAP
Input / output per 1M	$0.50 / $2.00	$0.50 / $2.00
Weight precision	8-bit	3-bit + expert pruning
Foundation size	Compact, dense	Large, sparse MoE
Vision	Yes	Yes
Context window	262,144	262,144
Schema · tools · streaming	Yes	Yes

Pick MiMo-V2.5 when the image is the hard part: dense OCR on imperfect scans, numeric extraction from charts and tables, screenshot and UI grounding, small-text reading, telling near-identical product photos apart — anything where a transposed digit is a defect rather than a cosmetic flaw. You want the highest-fidelity path from pixels to tokens, and 8-bit weights are how you get it at this price.

Pick MiniMax-M3 when the image is the easy part. A large sparse foundation compressed to 3 bits still holds more world knowledge and reasoning breadth than a compact dense model can, even after pruning. If the task is "look at this photo and reason at length about what it implies," or the image is one input in a mostly-text conversation, the bigger foundation is the better instrument — the visual acuity you traded away was not the binding constraint.

06MiMo-V2.5 vs MiMo-V2.5-Pro: the sibling that traded eyes for depth

This is the comparison that generates the most support tickets, so it is worth being blunt. MiMo-V2.5-Pro-W8A16 and MiMo-V2.5-W8A16 share a price, a context window, a precision tier, and a family name. Pro does not accept images. It is not a superset. It is a different allocation of the same budget.

Serving a multimodal model means paying for a vision tower and projector alongside the decoder — at a fixed cost target, capacity not spent on the language model itself. Pro spends all of it on text: a larger foundation with more headroom for long-form writing, extended reasoning, and code. MiMo-V2.5 spends part of it on sight.

So the decision rule is about your inputs, not your ambitions. If any request can contain a screenshot, a scanned page, a chart, a diagram, or a photograph, you need MiMo-V2.5 — Pro will reject the content block. If every request is text and you reached for the vision model out of habit, you are leaving foundation quality on the table for nothing: take MiMo-V2.5-Pro at the identical rate.

A pattern worth stealing: route by content type. Anything with an image part goes to MiMo-V2.5-W8A16, everything else to MiMo-V2.5-Pro-W8A16. Same family, same price, same context budget — the router is trivial and there is no cost model to reconcile.

07What "uncensored image understanding" actually means

There is no refusal layer in front of this model and no content filter on its output. That is a product decision with a specific consequence for vision work: the model describes what is in the frame instead of declining based on the subject matter.

This matters because filtered vision APIs are structurally unsuited to a whole category of legitimate work. A moderation pipeline must look at the material it moderates and say what it is; a classifier that refuses to examine the thing it classifies is not a classifier. The same holds across a long list of real jobs:

• Trust and safety tooling — labeling exactly the imagery a mainstream API will not process, which is the imagery that most needs labeling.
• Medical, forensic, and accident imagery — clinical, injury, and scene photography that generic filters reject as graphic regardless of who is asking.
• Security research — reading phishing pages, malicious interfaces, and captured screens without the model balking at the premise.
• Adult-content classification and archival cataloging — pipelines that exist to identify material and cannot outsource the looking, plus historical collections that filters treat as unlabelable.

Removing the refusal layer removes an engineering obstacle, not a legal one. Shannon AI is operated by Shannon Lab LLC in New Mexico, USA, and use of the API is governed by our Responsible Use Policy. Unfiltered output is a capability, not a permission slip.

08Calling it: endpoints, image input, and where it fits

The model is MiMo-V2.5-W8A16, reachable on three request dialects against the same weights, so you keep whichever SDK you already have:

• /v1/chat/completions — OpenAI-compatible. Images go in as image_url content parts, either as an https URL or a base64 data URI.
• /v1/messages — Anthropic-compatible. Images go in as image content blocks.
• /v1/responses — the Responses-style dialect.

Streaming and tool calling work on all three, so the model can look at a screenshot and call a function about what it saw in one turn. Full request and response schemas are in the API documentation, which also hosts a playground you can paste an image into before writing any code.

MiMo-V2.5 supports strict json_schema and response_format — the difference between a demo and a pipeline. Hand it a receipt plus a schema with typed fields for vendor, date, line items, and total, and you get back an object your code can validate rather than a paragraph you have to regex. Two of the twelve models in this catalog lack full structured-output support; this one does not, and that is why it works as a high-volume extraction service.

The context window is 262,144 tokens; generated output is capped at 16,384 tokens per request. Images consume input tokens as a function of resolution, so high-resolution page scans eat context far faster than an equivalent word count of text. Page long documents rather than loading a 300-page scan in one request, and treat the output cap as a real constraint on exhaustive extraction — chunk dense tables by row range instead of getting a truncated answer.

The reasoning_effort dial has real leverage here given the family's pedigree. Turn it down for high-volume classification, where the answer is a label and thinking tokens are pure cost; turn it up for chart interpretation, visual math, and diagram reasoning, where the model must combine several things it sees before it can answer.

Where it fits, and where it does not

Good fits. High-volume document and receipt extraction; OCR-adjacent work on imperfect scans; chart, dashboard, and table reading; screenshot understanding and UI grounding; catalog tagging and visual QA; moderation pipelines — any multimodal task where per-call cost binds and fine-detail precision is what you refuse to give up.

Poor fits. Frontier-difficulty multimodal reasoning needing the strongest model regardless of price — Kimi-K3's job, and it costs nearly ten times as much for a reason. Text-only generation, where MiMo-V2.5-Pro gives a better foundation at the same rate. Work leaning on encyclopedic world knowledge, where a large sparse foundation outruns a compact dense one. And anything needing more than 262K of context, which no model here offers.

09Frequently asked questions

---

Sources: MiMo technical report (arXiv 2505.07608) · MiMo-VL Technical Report (arXiv 2506.03569) · XiaomiMiMo/MiMo-VL · MiMo-VL-7B-RL-2508. All benchmark figures are Xiaomi-reported for the released checkpoints named above and describe the family lineage, not the exact build on this endpoint. Prices, context, and capability flags are Shannon AI Gateway list values, September 2026.