Uncensored MiniMax API: vision at the floor price

Almost every hosted multimodal API prices vision as a premium. You pay a text-model rate for text, and then you pay more — sometimes several times more — the moment an image enters the prompt. On top of that, image inputs are where refusal behavior is most aggressive: the model that will happily discuss a topic in text will decline to describe the same thing in a photograph. The result is a product that is expensive per call and unreliable per call, which is the worst of both worlds when you are processing volume. MiniMax-M3 on Shannon is our answer to that: the same $0.50 / $2.00 you pay for our cheapest text model, applied to a foundation that was trained multimodal from the first step, with no refusal layer sitting in front of it.

01What "cheapest uncensored vision model" actually means here

The Shannon catalog is twelve models. Five of them accept images. Their prices, per million tokens:

Vision-capable model	Input	Output	Structured output
MiniMax-M3 · 3BIT-REAP	$0.50	$2.00	json_schema + response_format
MiMo-V2.5 · W8A16	$0.50	$2.00	json_schema + response_format
Kimi-K2.6 · W4A16-AUTOROUND-REAP	$0.78	$3.67	json_schema + response_format
inkling · W4A16-AUTOROUND-REAP	$1.42	$6.07	response_format only
Kimi-K3 · 3BIT-REAP	$3.83	$19.12	json_schema + response_format

MiniMax-M3 sits at the floor — $0.50 / $2.00 — tied with MiMo-V2.5 and nothing below it, because $0.50 / $2.00 is the floor of our price list. Against the top of the vision range it is 7.7× cheaper on input and 9.6× cheaper on output than Kimi-K3.

Put that in a workload. Suppose you are triaging 1,000 images — screenshots, scanned pages, product photos — averaging roughly 1,200 tokens of image plus prompt going in and 300 tokens of answer coming out:

Shannon list prices, September 2026. Token counts are illustrative; your image tokenization depends on resolution and how many images you attach per call.

Two more reference points that are useful to keep in your head. Filling the entire 262,144-token context once costs about 13 cents. A maximum-length 16,384-token response costs about 3.3 cents. Those are the actual bounds of a single worst-case request, and they are small enough that the interesting constraint on a vision pipeline stops being money and starts being throughput.

02The MiniMax lineage: a lab that optimizes tokens, not headlines

MiniMax AI is a Shanghai lab whose entire open-weight track record is about one question: how do you make attention cheap enough that long context and multimodality stop being luxuries? Every release in the line is an answer to that question, and the answers build on each other.

MiniMax-01 — Lightning Attention, and vision from the start

The MiniMax-01 series (published as arXiv 2501.08313) introduced Lightning Attention, an I/O-aware implementation of a linear attention variant, inside a hybrid Mixture-of-Experts model of 456B total parameters with 45.9B active per token. The series shipped as a pair: MiniMax-Text-01 and MiniMax-VL-01, the vision model. VL-01 used the now-familiar ViT-MLP-LLM stack, but with a 303-million-parameter Vision Transformer trained from scratch on 694 million image-caption pairs rather than a borrowed encoder, then aligned and jointly fine-tuned across a multi-stage pipeline processing 512 billion tokens.

That detail matters for the article you are reading. Vision in this lineage is not a late adapter bolted onto a text model to check a box. MiniMax has been training its own visual tower since its first open release, which is why document understanding, UI and screenshot comprehension, and chart reading have been consistent strengths of the family rather than afterthoughts.

MiniMax-M1 — hybrid attention for reasoning at length

MiniMax-M1 (arXiv 2506.13585) was described by its authors as the first open-weight, large-scale hybrid-attention reasoning model. It kept the 456B / 45.9B MoE body and interleaved the two attention types: a conventional softmax-attention transformer block after every seven Lightning Attention blocks. The payoff was measured in compute rather than leaderboard position — MiniMax reported that at a 100K-token generation length, M1 consumed roughly 25% of the FLOPs of DeepSeek R1, while natively supporting a one-million-token context.

MiniMax-M2 — the small-activation pivot

MiniMax-M2 went the other direction on size and won attention for it: a 62-layer decoder with 229.9B total parameters but only ~9.8B active per token, using 256 fine-grained experts per MoE layer with sigmoid routing to eight of them, RoPE positions and QK-RMSNorm for training stability. It was explicitly an agentic-coding model — multi-file edits, run-fix loops, test-validated repairs — and it introduced the interleaved thinking format, where reasoning is carried between tool-call rounds instead of being discarded after each turn. M2 is the model that established MiniMax's reputation for delivering frontier-adjacent agentic behavior at a fraction of the per-token cost, purely by keeping activations tiny.

MiniMax-M3 — the foundation under this endpoint

MiniMax-M3 was released June 1, 2026, with weights on Hugging Face shortly after and a technical report at arXiv 2606.13392. It is a 428B-parameter Mixture-of-Experts model with roughly 23B active per token, a documented one-million-token context, and native image and video input — MiniMax describes it as mixed-modality trained "from the very first step," rather than text-pretrained and then multimodally adapted. On MiniMax's own published numbers it scores 80.5% on SWE-bench Verified and 59.0% on SWE-bench Pro. Those are vendor-reported figures on a coding benchmark, which is exactly how you should weigh them; we cite them because they are the lab's published claims, not because we have independently reproduced them.

03MiniMax Sparse Attention: why a vision model can be this cheap

The interesting engineering in M3 is MiniMax Sparse Attention (MSA), and it is the direct reason a multimodal model of this size can sit at our floor price. Images are token-expensive. A handful of high-resolution pages can put tens of thousands of tokens into a prompt before you have written a single instruction. Under standard attention, that cost compounds quadratically, and the model that reads twenty pages is not twice as expensive as the one that reads ten — it is much worse than twice.

MSA changes the shape of that curve. It is a block-sparse mechanism built on a Grouped Query Attention backbone, split into two branches:

• A lightweight Index Branch that scores key-value blocks — max-pooling scores across the tokens within each block — and selects the top-k per query position and per GQA group, always including the local block containing the query itself for training stability.
• A Main Branch that runs ordinary softmax attention only over the blocks the index selected.

The configuration in the MSA paper is k = 16 blocks of 128 tokens, which gives every query a fixed budget of 2,048 attended tokens. Fixed, meaning it does not grow with sequence length. That is the whole trick: as context grows, a dense baseline's attention cost keeps climbing while MSA's main-branch cost stays flat, so attention scales linearly instead of quadratically.

Figures from the MiniMax Sparse Attention report (arXiv 2606.13392): a 28.4× reduction in per-token attention FLOPs versus equivalently configured GQA at 1M context, and 14.2× prefill / 7.6× decode wall-clock speedups measured on H800 hardware in the paper's 109B-parameter study. MiniMax separately reports 9× prefill and 15× decode speedups for M3 versus M2 at 1M context. All are lab-reported.

The paper also does the check that matters for a vision article: on a natively multimodal 109B MoE, MSA held performance comparable to a full-attention baseline across image benchmarks including AI2D, ChartQA and MMMU, and across video benchmarks including EgoSchema, LongVideoBench and VideoMME — with the sparse variants sometimes improving on the longer-context video tasks. Sparsity here is not a quality tax you pay for speed; on long inputs it appears to be roughly free.

Why you should care as a customer: serving cost is what sets list price. A foundation whose attention cost stops growing is a foundation you can afford to point at a hundred thousand images. That is the mechanism behind "cheapest uncensored vision model" — it is not a promotional rate, it is an architecture.

04What 3BIT-REAP means, and where it costs you

A 428-billion-parameter model does not fit anywhere convenient at full precision. Our MiniMax-M3 build carries the 3BIT-REAP suffix, which describes two separate operations applied in a specific order.

REAP: delete experts before you compress them

REAP — Router-weighted Expert Activation Pruning — is a one-shot MoE compression method from Cerebras and the University of Calgary, published as arXiv 2510.13999 and accepted at ICLR 2026. Its argument is that for generative tasks, pruning experts beats merging them: merging destroys the router's fine-grained control and introduces an error you can never recover. So REAP instead scores each expert with a saliency criterion combining router gate values (how often and how strongly the router picks it) and expert activation norms (how much its output actually contributes), then deletes the least salient experts outright.

The authors report removing up to 50% of experts from models as large as one trillion parameters while largely preserving baseline quality — on Qwen3-480B-Coder-FP8 at 50% pruning, 97.6% of baseline non-agentic coding ability and 96.7% on agentic SWE-Bench were retained. Those are the paper's numbers on the paper's models, not ours on MiniMax-M3; we cite them so you can judge the method, not to imply an equivalent measurement here.

3-bit weights, second

The surviving weights are then stored at 3 bits. The ordering is deliberate and it is the point of the combination: prune first, quantize second. There is no sense in spending your precious three bits per weight faithfully encoding experts you are about to delete. Prune the sparse tail, then let the remaining budget go entirely to experts that actually fire.

The honest caveat

3BIT-REAP is the most aggressive tier we ship. We use it on the largest foundations — MiniMax-M3, Kimi-K3, DeepSeek-V4-Pro, GLM-5.2, Nemotron3Ultra — because it is the difference between those models being available at a sane price and not being available at all. Where a foundation is small enough not to need it, we do not use it: the W8A16 models in the catalog run 8-bit weights with 16-bit activations, and W4A16-AUTOROUND-REAP sits in between, using AutoRound's learned rounding to recover accuracy that naive round-to-nearest quantization gives away. Pick the tier that matches your tolerance. If you are extracting fields from ten thousand invoices, 3-bit is the correct trade. If you are doing one high-stakes forensic read of a single image, spend the money on Kimi-K3.

05What "no refusal layer" means when the input is an image

Uncensored is a word that gets used loosely, so here is the precise claim: there is no refusal layer and no output filtering on this endpoint. No classifier inspects your image before the model sees it. No policy model inspects the tokens on the way out. What the model produces is what you receive.

In text, the effect of that is familiar. In vision, it is different in kind, because refusal behavior on image inputs is where commercial multimodal APIs are at their most conservative and their most arbitrary. The characteristic failures look like this:

• Blanket identity refusals — declining to describe people in a photograph at all, including attributes that are plainly visible and entirely benign, because a policy cannot distinguish description from identification.
• Subject-matter bailouts — refusing to transcribe or summarize a document because of what the document is about, even though transcription is a mechanical task.
• Medical and forensic squeamishness — declining to describe clinical imagery, injury documentation, or accident photographs that a professional needs described.
• Security-tooling false positives — refusing to read a screenshot because it contains a terminal, a packet capture, or a debugger.
• Silent degradation — the worst one: not a refusal but a hedged non-answer, which passes your schema validation and poisons your dataset without ever raising an error.

The operational problem with all five is not that they are unfair. It is that they are unmeasurable. A wrong answer is a data point; you can score it, catch it in evaluation, and correct for it. A refusal is a hole. At batch scale, a 3% refusal rate distributed non-randomly across your inputs — concentrated, by construction, on exactly the hard and unusual cases — is a systematic bias in your results that no amount of downstream cleanup recovers. Removing the refusal layer removes a source of silent, correlated missing data.

Shannon is operated by Shannon Lab LLC, New Mexico, USA, under our Responsible Use Policy. Uncensored means the model does not editorialize about your inputs; it does not mean the account has no terms.

06The full specification

Field	MiniMax-M3 · 3BIT-REAP
Model id	MiniMax-M3-3BIT-REAP
Foundation	MiniMax-M3 (MiniMax AI)
Context window	262,144 tokens
Max output	16,384 tokens
Image input	Yes
Video input	Not exposed
Streaming	Yes
Tool calling	Yes
Reasoning	Yes (reasoning_effort)
Structured output	json_schema + response_format
Refusal layer	None
Endpoints	/v1/chat/completions · /v1/messages · /v1/responses
Served on	Shannon's own GPU cluster

Two lines in that table deserve emphasis. Video input is not exposed. The upstream M3 foundation was trained on video and MiniMax documents video input; our endpoint accepts images. Build against images. Second, the 262,144-token window is ours, not the upstream one-million figure — plan long-document and many-image workloads against 262,144 and you will never be surprised.

07Calling the uncensored MiniMax API

Three dialects, the same model behind all of them, streaming on all three. Use whichever your existing client already speaks; there is no capability difference between them. Point your base URL at Shannon and change the model id.

Image understanding via the OpenAI dialect

curl https://api.shannon-ai.com/v1/chat/completions \
  -H "Authorization: Bearer $SHANNON_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "MiniMax-M3-3BIT-REAP",
    "messages": [{
      "role": "user",
      "content": [
        { "type": "text", "text": "Transcribe every field in this scan. Do not summarize." },
        { "type": "image_url",
          "image_url": { "url": "data:image/png;base64,iVBORw0KGgo..." } }
      ]
    }],
    "max_tokens": 4096,
    "stream": true
  }'

Remote https:// URLs and base64 data URLs both work in the image_url part. Attach several image parts in one message when you need cross-page comparison — you have 262,144 tokens to spend and, at $0.50 per million, spending them is cheap.

Typed extraction with a JSON schema

Where this model earns its place in a pipeline is the combination of vision + strict schema + floor price. Not every model in our catalog has all three: Laguna-S-2.1 and inkling do not support json_schema. MiniMax-M3 does, which means you can get a validated object out of an image without a parsing layer:

{
  "model": "MiniMax-M3-3BIT-REAP",
  "messages": [ /* text part + image_url part */ ],
  "response_format": {
    "type": "json_schema",
    "json_schema": {
      "name": "invoice",
      "strict": true,
      "schema": {
        "type": "object",
        "properties": {
          "vendor":     { "type": "string" },
          "issued_on":  { "type": "string" },
          "total_cents":{ "type": "integer" },
          "line_items": { "type": "array", "items": { "type": "string" } }
        },
        "required": ["vendor", "issued_on", "total_cents", "line_items"],
        "additionalProperties": false
      }
    }
  }
}

Reasoning effort

The model exposes reasoning_effort. On bulk image classification, keep it low — you are asking a perceptual question, and reasoning tokens are output tokens you pay for. Raise it for the calls where the image is evidence in a longer argument: reading a diagram against a specification, reconciling a chart with a table, deciding whether two screenshots show the same defect.

The complete parameter reference, error semantics, and streaming formats for all three dialects live in the Shannon API documentation.

08Where MiniMax-M3 is the right call — and where it isn't

Reach for it when

• Volume image work. Dataset labeling, moderation research, catalog enrichment, screenshot triage, batch OCR-and-structure over scanned documents. The floor price is the entire argument.
• Images into typed records. Vision plus strict json_schema plus the cheapest tokens in the catalog is a combination only two models here offer.
• Long multimodal context. Many pages, or many frames, in a single request — the architecture underneath is built for exactly this shape of input.
• Research where refusals corrupt the sample. Any study whose validity depends on the model answering every item, including the uncomfortable ones.
• Agentic loops that read a screen. The M-series lineage is explicitly agentic, and tool calling is available.

Reach for something else when

• You need the highest vision ceiling and the budget exists. Kimi-K3 · 3BIT-REAP is 7.7× the input price for a reason; on a single high-stakes read, that is the right money to spend.
• You need maximum-fidelity weights. 3-bit is our most aggressive tier. The W8A16 models trade price for precision, and MiMo-V2.5 · W8A16 is vision-capable at the same $0.50 / $2.00 if fidelity per weight is what you are optimizing.
• You need video. We do not expose video input on this endpoint.
• The task is pure text at the frontier. DeepSeek-V4-Pro-0813 exists for that; you would be paying for a vision tower you never use.

Architecture, benchmark, and speedup figures for the MiniMax family are lab-reported by MiniMax AI; REAP figures are reported by the REAP authors on their own models and are not a measurement of this build. Shannon prices, context window, and capability flags are ours and are read from our live model catalog.

09Frequently asked questions

---

Sources: MiniMax-M3 model card · MiniMax-M1: Scaling Test-Time Compute Efficiently with Lightning Attention · MiniMax-01: Scaling Foundation Models with Lightning Attention · REAP the Experts: Why Pruning Prevails for One-Shot MoE Compression. MiniMax benchmark and speedup figures are lab-reported and provisional pending independent verification. Shannon prices and capability flags are current as of September 5, 2026.