# MiniMax M3.1 Flash MiniMax M3.1 Flash is the agent-oriented model MiniMax (the Chinese lab behind [[MiniMax Agent]]) released on 19 October 2026. The pitch is cost. MiniMax claims nearly 3× faster inference than the previous generation, 70% lower usage cost on agentic tasks, and scores close to the leading models on agent, coding, computer-use and cybersecurity benchmarks. Before launch it ran anonymously as "Space Bunny" and, according to MiniMax, processed more than 200 trillion tokens that way. Why does it matter? Because its price lands exactly on top of [[DeepSeek V4.1 Flash]]. If you run coding agents at volume, that's the comparison to make, and it's less flattering than the launch post suggests. ## What it is - **1M-token context**; text, image and video input - **Reasoning is always on**, with adjustable depth (MiniMax M3 let you switch it off) - **Throughput**: MiniMax claims 200+ tokens/s output, with a median (P50) above 170, which it credits to "more aggressive sparsity" - **Size**: the announcement doesn't say. One HN commenter quotes MiniMax's FAQ at 428B total and 23B active parameters, which would make it a [[AI Mixture of Experts (MoE)|mixture-of-experts]] model, nowhere near "small" - **Access**: MiniMax API, the M Plan subscription, and the MiniMax Code desktop app. The weights "will be released later", so for now it is NOT an [[AI Open Weight Models|open-weight]] model. A Pro model is announced for later as well ## Benchmarks (self-reported) | Benchmark | What it measures | M3.1 Flash | |---|---|---| | AutomationBench-AA | Enterprise workflows across SaaS tools | "among the top performers" (no number) | | BankerToolBench | Multi-step planning, financial compliance, chained tool calls | 76.2 | | DeepSWE v1.1 | Autonomous software tasks, validated by tests | 73.7% | | OSWorld V2 | Computer use across applications | 61.7 | | Agents' Last Exam (ALE) | Real professional work | 29.6 | | CyberGym | Vulnerability discovery and validation | 82% | The comparison I care about is the one MiniMax doesn't show. On the three benchmarks both labs publish, DeepSeek's own table for V4.1 Flash is higher: 74.2 on DeepSWE v1.1, 31.8 on Agent's Last Exam, 88.1 on CyberGym. MiniMax also calls its OSWorld V2 and ALE results the best among Chinese models, which doesn't square with DeepSeek's 31.8 on ALE. Different agent scaffolds or effort settings, probably. But that's exactly why two vendor tables can't be compared directly (one commenter also spotted that MiniMax's charts mix Opus 5 and Opus 5.5 from one graphic to the next). ## Pricing Flat, per million tokens: **$0.15 input, $0.60 output, $0.015 cached input**. minimaxir checked the "70% cheaper" claim on HN. The predecessor, MiniMax M3, cost $0.30 / $1.20 / $0.06. Input and output were halved, and cache reads went from 20% to 10% of the input price, so ~70% on cache-heavy agent work is plausible. Against DeepSeek V4.1 Flash it gets more interesting. Input and output match DeepSeek's off-peak rates, but DeepSeek charges $0.003 for a cache hit, five times less. DeepSeek doubles everything at peak hours; MiniMax stays flat. Since cache reads dominate agent loops, that one line decides the bill. Take the 447-turn coding task priced in my DeepSeek note (1.0M input, 165K output, 36.5M cache reads). By my arithmetic it costs about $0.80 on M3.1 Flash, $0.36 on DeepSeek off-peak, and $0.72 on DeepSeek at peak. Cheap everywhere, but not cheapest. ## Reception The [[Hacker News]] thread (512 points, 218 comments) was mostly skeptical: - **"Benchmaxed."** DCKing says Space Bunny was served for free as a stealth model in [[OpenCode]] Go and "failed to impress anyone"; people preferred GLM 5.3 Flash and DeepSeek V4.1 Flash. RussianCow's take on every MiniMax model so far: good on paper, disappointing in practice. XCSme's independent tests put it far below Opus 5 and GPT-5.6 Sol - **Counter-evidence exists.** chriddyp (Plotly) reports 41 of 43 on their data-analytics benchmark, tied with [[Claude Opus 5.5|Opus 5.5]] and [[Gemini 4 Argon|Argon]] at a third of the price. dom96 measured it on par with DeepSeek V4 Pro on his adversarial programming-language benchmark, one he doubts anyone tunes for - **Fast tokens, slow answers.** The same Plotly run made it the *slowest* model they tested, because it uses a LOT of tokens, not because of throughput. Between Space Bunny versions, time per question went from 30 seconds to 3 minutes while accuracy went up. LaurensBER's complaint fits: launches almost never show wall-clock time per task, and at work a "cheaper" model that takes twice as long is hard to justify - **Announcement fatigue.** Perenti and mawadev: every launch is on top of every chart and cheaper than everything, and nobody verifies it independently. jstummbillig's reply: people who care run their own evals, and a lab that posts silly numbers pays with its reputation - **What "Flash" means now.** Per polymorph1sm and penagwin, it now means "smaller than the 1T+ flagships", not small and not necessarily fast. terhechte hoped for a ~150B model that would run on a 128 GB Mac or a DGX Spark; if the 428B figure holds, that won't happen ## What I take from it - **Tokens per second isn't speed.** What you wait for is tokens per task divided by tokens per second, and a 200 tokens/s model that needs five times more tokens than a slower rival still loses. Ask for time per task, or measure it - **$0.15 / $0.60 is the new floor for agent-grade models**, and cache pricing is where providers now differ. Price your actual workload, cache reads included, before you switch - **Vendor benchmark tables are marketing**, a textbook case of [[Goodhart's Law]]. A stealth period like Space Bunny is a public blind test in all but name, and here the verdict was lukewarm. Run your own [[AI Evaluation|evals]] on your own tasks - **Wait for the weights** before deciding whether it's useful locally ## References - Announcement: https://www.minimax.io/blog/minimax-m31f - Product page and FAQ: https://agent.minimax.io/tools/m3-1-flash-preview - M Plan: https://www.minimax.io/m-plan - API: https://platform.minimax.io - Hacker News discussion (512 points): https://news.ycombinator.com/item?id=50154540 - Plotly data-analytics benchmark: https://plotly.com/blog/minimax-m3-1-flash-plotly-data-analytics-bench/ - AIBenchy comparison (XCSme): https://aibenchy.com/compare/minimax-minimax-m3-1-flash-high/deepseek-deepseek-v4-1-flash-1017-high/openai-gpt-5-6-sol-high/anthropic-claude-opus-5-high/ - Killswitch adversarial benchmark (dom96): https://bench.killswitch-lang.org/ ## Related - [[MiniMax Agent]] - [[DeepSeek V4.1 Flash]] - [[AI Mixture of Experts (MoE)]] - [[Sparse AI Models]] - [[AI Open Weight Models]] - [[Goodhart's Law]] - [[AI Evaluation]] - [[AI Inference]] - [[OpenCode]] - [[Claude Opus 5.5]] - [[Gemini 4 Argon]] - [[Hacker News]] - [[Large Language Models (LLMs)]]