DeepSeek vs MiniMax: Which Open-Weight MoE Wins in 2026?

If you are picking between two Chinese AI labs’ open-weight models for a coding agent or a high-volume API workload, the deepseek vs minimax decision usually comes down to four things: cost per million tokens, coding benchmarks, context window, and licence terms. Both lineups have moved on since the spring. DeepSeek’s API now serves deepseek-flash (DeepSeek-V4.1-Flash-1017, released 17 October 2026) and deepseek-v4-pro (DeepSeek-V4-Pro-0813, released 13 August 2026). MiniMax’s API platform now lists two language models, MiniMax-M3.1-Flash and MiniMax-M3, and says its M2.5 and M2.7 families stopped accepting new users on 19 October 2026 and will be retired on 18 November 2026. This article gives you the numbers, a worked cost example and a pick for each scenario.

The verdict up front

At the budget tier the two are now neck and neck. MiniMax-M3.1-Flash lists $0.15 input / $0.60 output per 1M tokens at all hours; deepseek-flash lists $0.30 / $1.20 at peak and exactly $0.15 / $0.60 off-peak, with much cheaper cache hits ($0.006 peak / $0.003 off-peak against MiniMax’s $0.015). So MiniMax is cheaper if your traffic runs in DeepSeek’s peak hours; DeepSeek is slightly cheaper off-peak and on cache-heavy prompts. Both have a 1M-token context and accept images; MiniMax-M3.1-Flash also accepts video.

The bigger differences are elsewhere. Pick DeepSeek if you want MIT-licensed weights with no commercial conditions, a non-thinking mode for latency-sensitive calls, and a second, larger model (DeepSeek V4-Pro) on the same API. Pick MiniMax if you want video input, MiniMax’s coding-tool plans (it documents setups for Claude Code, Codex, Cursor, OpenCode and others), or a platform whose privacy policy says personal data is stored in the United States rather than China.

At-a-glance comparison

Figures from DeepSeek’s pricing page, MiniMax’s pricing overview and model list, and the model cards, checked 26 October 2026. Prices are USD per 1M tokens.

MiniMax shows its MiniMax-M3 standard prices as a “Permanent 50% off” the original $0.60 / $2.40, and charges double ($0.60 / $2.40) for requests over 512K input tokens; a Priority tier costs 1.5× standard. DeepSeek’s rates do not change with prompt length. DeepSeek’s maximum output is 384,000 tokens on both models.

Coding

This is where the comparison gets interesting, and MiniMax publishes the head-to-head. The MiniMax-M3.1-Flash model card compares it directly with DeepSeek-V4.1-Flash-1017 (MiniMax-reported; the DeepSeek coding figures match DeepSeek’s own V4.1-Flash-1017 card):

On MiniMax’s own table, DeepSeek-V4.1-Flash-1017 leads the three coding benchmarks and MiniMax-M3.1-Flash leads Automation-Bench and, narrowly, HLE. DeepSeek’s own card lists 42.6 on full HLE without tools and 44.0 on the text-only subset, so that row is not like-for-like. MiniMax also reports results where it gives no DeepSeek figure, such as 82.2 on CyberGym and 76.2 on BankerToolBench. MiniMax lists mini-SWE-agent for DeepSWE v1.1 and Codex for Terminal Bench 4.0, while DeepSeek’s card says its figures used the Mini-SWE-V2 and Claude Code harnesses; scores in other harnesses will differ.

For coding workflows, see the DeepSeek for coding walkthrough or the DeepSeek Coder vs Copilot comparison.

Practical coding pick

• Coding benchmarks on the budget tier: deepseek-flash, which leads MiniMax-M3.1-Flash on DeepSWE v1.1, Terminal Bench 4.0 and NL2Repo-Bench in MiniMax’s own table.
• Agentic loops in coding tools: either. DeepSeek’s API docs list integration guides for tools such as Claude Code, OpenCode and GitHub Copilot; MiniMax documents its M Plan for MiniMax Code, Claude Code, Codex, Cursor, OpenCode and others. MiniMax recommends temperature=1.0 and top_p=0.95 for M3.1-Flash.
• Cost-sensitive bulk code review: MiniMax-M3.1-Flash during DeepSeek’s peak hours; deepseek-flash off-peak, especially with long cached prefixes.

Reasoning and thinking modes

Both labs ship reasoning capability, and the API ergonomics are now similar.

DeepSeek puts thinking behind request parameters on both models: it is on by default (effort high), you can set reasoning_effort to "low", "high" or "max", and you can opt out with extra_body={"thinking": {"type": "disabled"}}. The response returns reasoning_content alongside the final content. No separate model ID, no version juggling.

MiniMax’s model-invocation guide says MiniMax-M3.1-Flash thinks before it answers by default, with an effort setting of low, medium, high, xhigh or max (default max). On its OpenAI-compatible protocol, thinking comes back in reasoning_content and content holds only the final answer; on its Anthropic-compatible protocol it comes back as thinking content blocks. MiniMax’s guide says thinking cannot be turned off on MiniMax-M3.1-Flash (sending thinking: {"type": "disabled"} returns a 400 error) and suggests lowering effort instead; its API reference says MiniMax-M3 skips thinking when it is disabled.

For more on DeepSeek’s reasoning lineage, the DeepSeek R1 page covers the original step-by-step reasoning release.

Writing and general chat

For long-form writing, customer support, summarisation and translation, the deciding factors are price, context and output length:

• Cost. MiniMax-M3.1-Flash’s $0.15 / $0.60 matches deepseek-flash’s off-peak rate at all hours; at peak DeepSeek costs twice as much.
• Context. Both now accept 1M tokens, enough for whole-document summarisation or long RAG passes; see building a DeepSeek RAG pipeline for setup.
• Output ceiling. DeepSeek allows up to 384K output tokens in one call, which helps with very long drafts.

If you skip thinking for short chat replies, DeepSeek’s non-thinking mode keeps output tokens, and so cost, down. If your team runs content creation workflows, test both on your own prompts.

Pricing — worked example

Cost comparisons fail when articles quote a single rate. The honest version enumerates all three token buckets. Here is a representative workload: 1,000,000 API calls, each with a 2,000-token system prompt (cached across calls), a 200-token user message (uncached) and a 300-token response, which is 2,000M cached input tokens, 200M uncached input tokens and 300M output tokens.

DeepSeek deepseek-flash

                   PEAK                                OFF-PEAK
Cached input    : 2,000M × $0.006/M = $12.00           × $0.003/M = $6.00
Uncached input  :   200M × $0.30/M  = $60.00           × $0.15/M  = $30.00
Output          :   300M × $1.20/M  = $360.00          × $0.60/M  = $180.00
                                      -------                       -------
Total                                 $432.00                       $216.00

MiniMax-M3.1-Flash and MiniMax-M3 (MiniMax pricing page)

MiniMax-M3.1-Flash
Cached input    : 2,000M × $0.015/M = $30.00
Uncached input  :   200M × $0.15/M  = $30.00
Output          :   300M × $0.60/M  = $180.00
                                      -------
Total                                 $240.00

MiniMax-M3 (standard, up to 512K input)
Cached input    : 2,000M × $0.06/M  = $120.00
Uncached input  :   200M × $0.30/M  = $60.00
Output          :   300M × $1.20/M  = $360.00
                                      -------
Total                                 $540.00

DeepSeek deepseek-v4-pro

                   PEAK                                OFF-PEAK
Cached input    : 2,000M × $0.044/M = $88.00           × $0.022/M = $44.00
Uncached input  :   200M × $1.32/M  = $264.00          × $0.66/M  = $132.00
Output          :   300M × $3.96/M  = $1,188.00        × $1.98/M  = $594.00
                                      ---------                     -------
Total                                 $1,540.00                     $770.00

For this workload, deepseek-flash run entirely at peak ($432) costs 1.8× MiniMax-M3.1-Flash ($240), while deepseek-flash run entirely off-peak ($216) is about 10% cheaper than MiniMax-M3.1-Flash. deepseek-v4-pro costs about 2.9× MiniMax-M3 at peak ($1,540 vs $540) and about 1.4× off-peak ($770). A real workload usually straddles DeepSeek’s two periods, so your DeepSeek bill lands between those totals. Use the DeepSeek pricing calculator to plug in your own ratios.

API access and developer ergonomics

Both providers expose OpenAI-compatible HTTP APIs. DeepSeek’s chat requests go to POST /chat/completions at https://api.deepseek.com; DeepSeek also accepts the Anthropic API format at https://api.deepseek.com/anthropic and the OpenAI Responses API format. The API is stateless: clients must resend the full conversation history with every request.

A minimal DeepSeek call in Python:

import os
from openai import OpenAI

client = OpenAI(
    api_key=os.environ.get("DEEPSEEK_API_KEY"),
    base_url="https://api.deepseek.com",
)

resp = client.chat.completions.create(
    model="deepseek-flash",
    messages=[{"role": "user", "content": "Summarise the migration plan."}],
    reasoning_effort="high",
    extra_body={"thinking": {"type": "enabled"}},
    max_tokens=1024,
)
print(resp.choices[0].message.reasoning_content)  # reasoning_content
print(resp.choices[0].message.content)            # the final answer

To skip thinking for a quick reply, pass extra_body={"thinking": {"type": "disabled"}} instead. See the DeepSeek OpenAI SDK compatibility notes for the full surface.

MiniMax’s API docs give an OpenAI-compatible base URL of https://api.minimax.io/v1 and an Anthropic-compatible one of https://api.minimax.io/anthropic, and also document an OpenAI Responses-compatible endpoint. For self-hosting, the MiniMax-M3.1-Flash card recommends SGLang, vLLM and TokenSpeed, and includes a DSpark speculative-decoding draft head. DeepSeek’s current weights are large too: about 893 GB for DeepSeek-V4-Pro-0813 and about 510 GB for DeepSeek-V4.1-Flash-1017 on Hugging Face.

Developer features worth comparing:

• JSON mode: DeepSeek supports it; it is designed (not guaranteed) to return valid JSON, with the documented caveat that prompts include the word “json” plus an example schema and that max_tokens is set high enough to avoid truncation.
• Tool calling / function calling: both providers support it in OpenAI-compatible format.
• Streaming: both support SSE, and both return thinking in a separate reasoning_content field on the OpenAI-compatible protocol.
• Context caching: both bill cache hits at a lower rate. On deepseek-flash a cache hit costs 2% of the cache-miss rate; on MiniMax-M3.1-Flash a cache read costs 10% of the input rate.
• FIM completion (Beta): DeepSeek supports it in non-thinking mode only.

Privacy and ecosystem

The two platforms handle data differently. DeepSeek’s Privacy Policy (last updated 10 February 2026) says it collects, processes and stores personal data in the People’s Republic of China; see the DeepSeek privacy guide for a longer treatment. MiniMax’s API platform privacy policy names Nanonoble Pte. Ltd. in Singapore as the operator and says personal data is stored in a data centre in the United States. The strongest mitigation is the same for both: self-host the open weights.

Licences differ more than the old “MIT vs Modified-MIT” framing suggests. DeepSeek-V4.1-Flash-1017 and DeepSeek-V4-Pro-0813 ship under MIT for both code and weights. MiniMax-M3.1-Flash and MiniMax-M3 ship under the MiniMax Community License, which grants rights for non-commercial use; commercial use requires displaying “Built with MiniMax M3.1-Flash” (“Built with MiniMax M3” for M3) and notifying MiniMax, with prior written authorisation above US$20 million in yearly revenue, and the licence lists prohibited uses, including military purposes. Read the licence before you commit.

Ecosystem-wise, both lean on coding tools. DeepSeek’s API docs list guides for tools such as Claude Code, Codex, OpenCode and GitHub Copilot. MiniMax sells subscription plans (M Plan and Token Plan) with setup guides for MiniMax Code, Claude Code, Codex, Cursor, OpenCode, OpenClaw and others, and says M Plan and Token Plan requests to the retiring M2.5 and M2.7 names are now routed to M3.1-Flash.

When to pick which

A clear decision rule beats a benchmark table:

• Pick deepseek-flash if your workload can run off-peak or reuses long cached prefixes, you want the stronger coding scores in MiniMax’s own comparison, or you need MIT-licensed weights.
• Pick deepseek-v4-pro if you want DeepSeek’s second current model with the 1M context and up to 384K output, at $3.96 / $1.98 per 1M output tokens (peak / off-peak).
• Pick MiniMax-M3.1-Flash if your traffic is concentrated in DeepSeek’s peak hours, you need video input, or you want MiniMax’s coding-tool plans.
• Pick MiniMax-M3 if you need a MiniMax model with a non-thinking mode or the Priority tier, or have a reason to stay on the earlier M3 release; at $0.30 / $1.20 it costs more than M3.1-Flash, which MiniMax builds on M3.

Alternatives worth considering

Neither model is the only Chinese open-weight option. DeepSeek vs Qwen covers Alibaba’s family; DeepSeek vs Kimi compares against Moonshot AI’s Kimi K3 line. For a wider survey, see our best Chinese AI models roundup and the broader AI comparison hub.