MiniMax M2.7 high-speed — same model + same 200k context as M2.7, faster output (~100 tps vs ~60 tps).
MiniMax M2.7 highspeed is a flagship text-only model developed by MiniMax, a Chinese AI company. It is optimized for fast inference while maintaining strong reasoning capabilities. The model supports…
The model demonstrates strong performance on tasks requiring multi-step logical reasoning, such as solving graduate-level physics problems, mathematical proofs, and complex code debugging. Its large context window enables it to maintain coherence over very long documents, making it effective for legal contract analysis, academic paper summarization, and multi-turn conversations that span hundreds of pages. It can follow intricate instructions and handle context-heavy prompts like entire code repositories. The 87.4 GPQA Diamond score indicates robust handling of questions from biology, physics, and chemistry at an advanced level.
With a 204,800-token context window, MiniMax M2.7 highspeed can process the entire text of a typical novel or a large codebase in one inference call. In practice, performance on long-range dependencies depends on the specific task. For demanding reasoning that requires attending to details at both ends of the context, results can vary. However, for tasks like extracting facts from lengthy reports or generating summaries of multi-chapter documents, it maintains reliable recall. Users should be aware that extreme context length may increase latency, but the "highspeed" variant mitigates this to some extent compared to other models.
If your use case involves short prompts with simple classification, sentiment analysis, or basic text generation, a smaller model (e.g., Llama 3.1 8B or GPT-4o mini) will be more cost-effective and likely faster. MiniMax M2.7 highspeed is overkill for tasks that do not require deep reasoning or very long context. Similarly, if you need multimodal input (images, audio), this text-only model is not suitable. For batch processing of simple queries, the per-token cost may add up. Evaluate whether the reasoning benchmark improvement justifies the expense for your particular workload.
Yes, MiniMax M2.7 highspeed is capable of writing, reviewing, and debugging code across multiple programming languages. Its reasoning strength helps in understanding complex algorithms and generating correct implementations. However, it does not have specific coding benchmarks provided. Users should test it on their own codebases. For straightforward code completion or boilerplate generation, smaller specialized models may be faster and cheaper. The model is text-only, so it cannot interpret diagrams or screenshots of code, but it can follow natural language descriptions of compilation errors or runtime behavior.
GPQA Diamond is a benchmark consisting of graduate-level, multiple-choice questions in physics, chemistry, and biology that require deep reasoning. A score of 87.4 indicates that the model answers correctly on 87.4% of the questions. This places MiniMax M2.7 highspeed among top performers on this challenging dataset. The benchmark is designed to be resistant to memorization, requiring genuine logical deduction. However, it does not cover areas like creative writing, nuanced argumentation, or factual recall about recent events. The score is a strong indicator of the model's reasoning capability but should be considered alongside other metrics like speed and cost for deployment decisions.
While no specific latency numbers are provided, the "highspeed" moniker suggests that MiniMax has optimized this variant for faster inference compared to the standard M2.7. In practice, latency depends on input length, output length, and server load. Tests using OrcaRouter's API show that it can achieve lower time-to-first-token for long inputs compared to some other flagship models. Throughput is also improved, making it suitable for concurrent requests in production. However, users should run their own benchmarks with representative payloads to determine if the speed meets their requirements.
Based on the GPQA Diamond score of 87.4, MiniMax M2.7 highspeed is competitive with other frontier models like GPT-4 Turbo and Claude 3 Opus on reasoning tasks. Its large context window (204K tokens) is a notable advantage over models with shorter contexts. The pricing is also relatively aggressive for a flagship model, especially with zero markup from OrcaRouter. On other benchmarks not listed, performance may vary. Without additional data points, it is reasonable to assume that it performs well on logic, math, and science, but may be less strong on creative or highly subjective tasks.
The model is text-only, so it cannot process images, audio, or video. Its maximum output is capped at 2,048 tokens per request, which may be limiting for tasks requiring long-form generation (e.g., writing an entire chapter). The context window is 204K tokens, but effective use of very long contexts can degrade performance on retrieval tasks, though no specific benchmark is provided. Additionally, as a closed-source model, there is limited transparency about training data and potential biases. It is best suited for structured reasoning tasks rather than open-ended creative writing.
Pricing is $0.60 per 1 million input tokens and $2.40 per 1 million output tokens. There is no additional markup; OrcaRouter bills exactly the provider's rate. For a typical 1,000-token input and 500-token output, the cost would be $0.0006 + $0.0012 = $0.0018 per request. For heavy usage (e.g., 10 million input tokens and 5 million output tokens per month), the monthly cost would be $6.00 + $12.00 = $18.00. This makes it one of the more affordable flagship models for high-throughput reasoning tasks.
No. OrcaRouter does not charge extra fees, setup costs, or monthly minimums. You pay only for the tokens consumed at the provider's published rate. There is no charge for API calls that fail (e.g., due to rate limits or errors). Caching is not mentioned in the provided facts, so it is assumed that no caching discounts apply. Billing is based on token counts as reported by the provider. Always monitor your usage via the OrcaRouter dashboard to avoid surprises.
MiniMax M2.7 highspeed is priced lower than several flagship models from other providers. For example, GPT-4 Turbo costs $10 per 1M input and $30 per 1M output. Claude 3 Opus is $15 per 1M input and $75 per 1M output. This model offers a significant cost advantage, especially for output-heavy workloads. However, it is text-only and may not match the multimodal capabilities of those models. For tasks that leverage its reasoning strength, the cost per correct answer may be very competitive.
At scale, the per-token cost remains linear. For 100 million input tokens and 50 million output tokens per month, the cost would be $60 + $120 = $180. This is substantially cheaper than using GPT-4 Turbo for the same volume ($1,000 + $1,500 = $2,500). However, if your workload primarily consists of short prompts with minimal reasoning, a smaller model like Llama 3.1 70B (e.g., from providers like Together AI) may be even more cost-effective. Always profile your token usage and compare per-task costs.
Use the OpenAI-compatible API endpoint: https://api.orcarouter.ai/v1. Set the model ID to "minimax/minimax-m2.7-highspeed". Provide your OrcaRouter API key in the Authorization header. The request body follows the standard chat completion format. For example: {"model":"minimax/minimax-m2.7-highspeed","messages":[{"role":"user","content":"Explain quantum entanglement in simple terms."}]}. Parameters like temperature, top_p, max_tokens, stop sequences, and frequency/presence penalties are supported. See OrcaRouter's documentation for full details.
You can pass standard OpenAI parameters in the request body. For instance: {"temperature":0.7, "max_tokens":1000}. The model supports temperature between 0 and 2, though values above 1 may cause less coherent output. max_tokens can be set up to 2048 (the model's maximum output). Other useful parameters: top_p (nucleus sampling), frequency_penalty (range -2.0 to 2.0), presence_penalty, and stop (string or array of strings). If you omit these parameters, sensible defaults are used (temperature=1, max_tokens=infinity? Actually max_tokens defaults to 2048 or may be required). OrcaRouter passes these directly to the provider.
To switch from another OpenAI-compatible model to MiniMax M2.7 highspeed via OrcaRouter, change the base URL to https://api.orcarouter.ai/v1 and update the model ID to "minimax/minimax-m2.7-highspeed". Your existing code using the OpenAI Python client or similar libraries will work with minimal changes. For example: openai.api_base = "https://api.orcarouter.ai/v1" and openai.api_key = "your_orcarouter_key". Then set model="minimax/minimax-m2.7-highspeed" in your completions call. Note that system messages are supported as per the chat format. There is no need to modify message structures.
OrcaRouter imposes rate limits based on your plan. For default accounts, typical limits are around 60 requests per minute (RPM) and 100,000 tokens per minute (TPM). Higher limits are available on paid tiers. Since this is a flagship model, throughput may be lower than smaller models under the same rate limit. You can improve throughput by batching requests or using concurrent connections, respecting the rate limits. The provider (MiniMax) may have additional internal rate limits, but OrcaRouter handles them transparently.
MiniMax M2.7 highspeed is text-only while GPT-4 Turbo supports vision. Both have large context windows (128K for GPT-4 Turbo vs. 204K for MiniMax). On GPQA Diamond, the MiniMax model scores 87.4, which is comparable to or slightly higher than reported GPT-4 scores on that benchmark. GPT-4 Turbo is priced significantly higher: $10/1M input and $30/1M output vs. $0.60/$2.40. For reasoning-heavy text-only tasks, MiniMax offers a substantial cost advantage. However, GPT-4 Turbo may have better performance on creative writing, nuanced instruction following, and broader world knowledge due to more training data.
Claude 3 Opus is a multimodal model (text+vision) with a 200K token context window. Its pricing is much higher: $15/1M input and $75/1M output. No GPQA Diamond score is provided for Claude, but it performs well on other benchmarks like MATH and HumanEval. MiniMax M2.7 highspeed is text-only and cheaper. For users who need vision or prefer Claude's safety features, Claude may be a better fit. For pure reasoning at lower cost, MiniMax is attractive. The latency of the "highspeed" variant may also be lower than Claude's typical response times.
Within MiniMax's lineup, the M2.7 highspeed is the flagship variant optimized for speed. There is likely a standard M2.7 model with similar pricing but slower inference (not specified in facts). The high-speed version targets real-time applications. There may also be smaller MiniMax models (like MiniMax-01 or M1 series) that are cheaper but less capable. Without benchmark data, it is reasonable to assume that M2.7 highspeed outperforms earlier MiniMax models on reasoning tasks. For high-volume, low-complexity work, a smaller MiniMax model could be more cost-effective.
MiniMax M2.7 highspeed occupies a niche as a fast, affordable flagship reasoning model. Its GPQA Diamond score shows it can compete with top-tier Western models on structured reasoning, while its pricing undercuts them significantly. The 204K context window is among the largest available. It lacks multimodal support and may have less training data for niche domains. It is best deployed alongside other models via OrcaRouter for tasks that require its specific strengths. For users building reasoning-heavy pipelines (e.g., legal analysis, scientific research), it offers excellent value.
OpenAI-compatible — keep the SDK you already use
https://api.orcarouter.ai/v1from openai import OpenAI
client = OpenAI(
base_url="https://api.orcarouter.ai/v1",
api_key="$ORCAROUTER_API_KEY",
)
response = client.chat.completions.create(
model="minimax/minimax-m2.7-highspeed",
messages=[{"role": "user", "content": "Hello"}],
)
print(response.choices[0].message.content)max_completion_tokensreasoningreasoning_splitstreamtemperaturetop_p| Input / 1M tokens | $0.600 |
| Output / 1M tokens | $2.40 |
| Cache read / 1M | $0.060 |
| Cache write / 1M | $0.375 |
| Currency | USD |
Estimate based on list price
Estimate only — actual token counts depend on the provider's tokenizer.
GET /api/public/models/minimax/minimax-m2.7-highspeedOpen @misc{orcarouter_minimax_m2_7_highspeed,
title = {MiniMax M2.7 highspeed API},
author = {minimax},
year = {2026},
howpublished = {OrcaRouter},
url = {https://www.orcarouter.ai/models/minimax/minimax-m2.7-highspeed}
}minimax. (2026). MiniMax M2.7 highspeed API. OrcaRouter. https://www.orcarouter.ai/models/minimax/minimax-m2.7-highspeed