ERA OF MINDS

Is MiniMax M3.1 Flash still free? What changed on October 22 and 23, and what it costs now

By Era of Minds ·

Short answer

Mostly no. The MiniMax M3.1 Flash free launch period ended on October 22, 2026: OpenRouter retired the :free model at 12:00 UTC and Ollama Cloud stopped at 23:59 UTC. The API now costs $0.15 per million input tokens and $0.60 per million output. On October 23 MiniMax published the weights on Hugging Face, free for non-commercial use, but they are about 250 GB and need 4 to 8 data-center GPUs. For most people, "free" now means a few MiniMax Agent credits or switching to another free model.

Era of Minds cover: MiniMax M3.1 Flash, still free? Not since Oct 22. Four free routes marked ended (OpenRouter, Vercel AI Gateway, Ollama Cloud, OpenCode Zen), the paid API price and the open-weights facts, checked October 23, 2026.
Figure: our own summary of the free routes, prices and weights, from the OpenRouter, Vercel, Ollama, OpenCode, Hugging Face and vLLM pages, checked October 23, 2026.

Is MiniMax M3.1 Flash still free?

Not through the routes most people used last week. MiniMax launched M3.1 Flash on October 19, 2026, and for about three days you could point Claude Code, Codex or OpenCode at it for $0. That window is closed. Here is what we found when we checked each route on the morning of October 23 (Dubai time):

  • OpenRouter: the free model page now says "The free launch period ended Oct 22, 2026." Before that it said the free period ends Thursday, October 22 at 12:00 UTC, 72 hours after launch. OpenRouter's public model list no longer has a :free entry for M3.1 Flash.
  • Vercel AI Gateway: the October 19 changelog said free "until Oct 22."
  • Ollama Cloud: free until 23:59 UTC on October 22, per Ollama's announcement. The model page now shows $0.15 input, $0.015 cached and $0.60 output per million tokens.
  • OpenCode Zen: it was listed as minimax-m3.1-flash-free during launch week. On October 23 OpenCode's model list has no M3.1 Flash entry at all, free or paid.

So if your agent started throwing errors or billing you this morning, that is why. Swap the :free slug for the paid one or pick another model.

Why was everyone talking about MiniMax M3.1 Flash?

Price and speed, plus a good backstory. MiniMax's launch post on October 19 revealed that "Space Bunny," an anonymous model that had been topping usage charts, was an early version of M3.1 Flash. MiniMax's release blog says Space Bunny processed more than 200 trillion tokens before anyone knew whose it was.

When we checked through the X API on October 23, the launch post had 1,276,369 views, 7,429 likes and 1,441 bookmarks. The weights announcement, posted at 01:06 UTC on October 23, had 108,052 views and 1,538 likes within about five hours.

The post that sold developers on it came from Hassan (@nutlope, who also builds the Hallmark design skill at Together AI; 134,796 views, 1,436 likes): he asked M3.1 Flash and Claude Opus 5.5 to build the same landing page and reported a cost of $0.014 against $1.22, or 88x cheaper, while running 4x faster. Others posted one-prompt Three.js scenes, like @NickDevFE's room-from-a-photo demo (14,701 views, 471 bookmarks). Those are single tests by the people posting them, not benchmarks, but they explain the rush.

How much does MiniMax M3.1 Flash cost now?

MiniMax's pricing page lists a flat rate for MiniMax-M3.1-Flash:

  • Input: $0.15 per 1M tokens
  • Output: $0.60 per 1M tokens
  • Cached input: $0.015 per 1M tokens

On OpenRouter four hosts served it on October 23. Parasail, Together and MiniMax charged the same list price, and inference.net was cheaper at $0.10 input and $0.55 output, though it was also the slowest at 22 tokens per second against Parasail's 126.

What that means in practice: a coding-agent task that reads 1 million tokens of your repo and writes 100,000 tokens costs about $0.21 at MiniMax's rate, before caching. Agents re-read the same files a lot, so cached input at a tenth of the price brings real bills lower. If token spend is what pushed you to free models in the first place, a compression proxy like Headroom for Claude Code attacks the same problem from the other side.

The cost-per-1M number is low, but this model always thinks before it answers. You can set the effort from low to max, and output tokens include that reasoning, so a cheap rate does not mean a tiny bill on long agent runs.

Where can you still use MiniMax M3.1 Flash for free?

We found three options on October 23. None of them is the unlimited free API of launch week.

  1. MiniMax Agent credits. MiniMax's agent terms say free users "receive free credits" on registration. A Korean reviewer at itplain found the web app labels the model "M3.1-Flash (Free)," but a new account was blocked with "No active plan available" until it claimed a 400-credit daily check-in. Good for trying it in a browser, not for wiring into an agent.
  2. AnyRouter. This smaller third-party router announced on October 20 that M3.1 Flash is free on every plan with no end date, and its public model list still showed $0 input and $0 output on October 23. It is not a big-name provider, so read its privacy and data terms before you send it private code.
  3. The open weights. Free to download and free for non-commercial use. The catch is hardware, covered below.

If what you actually want is a free model for your coding agent, OpenCode's Zen list on October 23 still had several marked free "for a limited time," including deepseek-v4.1-flash-free and glm-5.3-flash-free, both served by providers with a zero-retention policy. DeepSeek V4.1 Flash is the model MiniMax compares itself against in its own benchmark table. Our roundup of AI tools for coding covers the agents you would plug them into.

Can you download the MiniMax M3.1 Flash weights?

Yes, since October 23. MiniMax had promised the date in its launch post ("Model weights will arrive on Oct 23") and kept it. The MiniMaxAI/MiniMax-M3.1-Flash repository on Hugging Face is public and not gated, with a mirror on ModelScope.

We listed the repo files through the Hugging Face API:

  • 48 safetensors shards totalling 250.1 GB (232.9 GiB) for the main model
  • a separate dspark/ folder with a 2.3 GB draft head for speculative decoding
  • tags for mxfp8 and 8-bit: the model card says it was quantization-aware trained, and the vLLM recipe describes MXFP8 dense layers with NVFP4 experts

The model card lists about 428B total parameters, 23B active per token, a context window up to 1,000,000 tokens, and text, image and video input with text output. The download command is hf download MiniMaxAI/MiniMax-M3.1-Flash --local-dir MiniMax-M3.1-Flash, and at 250 GB you will want a fast connection and a big disk.

Can you run MiniMax M3.1 Flash locally?

Only if "locally" means a GPU server. The vLLM recipe (vLLM 0.33.0 or newer) runs it three ways:

  • 4 GPUs (tensor parallel 4): NVIDIA Blackwell (B200, GB200, B300, GB300) or AMD MI355X
  • 8 GPUs (tensor parallel 8): NVIDIA H100 or H200, with FP8 caches, a 262,144-token context in the example config, video input switched off and no speculative decoding
  • Two nodes with prefill and decode split, for high throughput

An 8x H100 node is the floor here, so this is not a Mac Studio or RTX 4090 model. Ollama confused a lot of people: its library page has only one tag, minimax-m3.1-flash:cloud. Running ollama run minimax-m3.1-flash:cloud sends your prompts to Ollama's servers and bills your account; nothing downloads to your machine. The model card says Transformers, KTransformers and Unsloth support is "in progress," so smaller quantized builds may come later. We would not count on one that fits consumer hardware soon at 428B parameters.

Is MiniMax M3.1 Flash open source?

Open weights, yes. Open source in the usual sense, no. The weights ship under the MiniMax Community License, which we read in full:

  • Non-commercial use is free: use, copy, modify, publish and redistribute.
  • Commercial use means you must "prominently display 'Built with MiniMax M3.1-Flash'" on a related website, UI, blog post or docs, and send MiniMax a one-time notice email.
  • Over $20 million a year in revenue: you need "separate, prior written authorization" before commercial use.
  • Commercial use explicitly includes selling API access to the model and deploying fine-tuned versions.
  • An appendix bans some uses outright, including "any military purpose."

That is closer to a source-available license than to Apache-2.0 or MIT. It is the same distinction we made yesterday with Supabase, where the code is genuinely open but the service is not. Here the weights are open, but your rights depend on what you build and how big you are.

How good is MiniMax M3.1 Flash, really?

Strong for the price, with one real weakness. MiniMax's own table on the model card puts it close to Claude Opus 5 on some tests and well behind on others:

  • DeepSWE v1.1 (coding): 73.7% vs Opus 5 74.0% and DeepSeek V4.1 Flash-1017 77.0%
  • Terminal Bench 4.0: 38.1 vs Opus 5 51.8
  • OSWorld 2.0 (computer use): 61.7 vs Opus 5 70.6
  • BankerToolBench: 76.2, tied with Opus 5

Those are the vendor's numbers. The independent ones, from Artificial Analysis as shown on OpenRouter on October 23: an Intelligence Index of 42.7, ranked #21 of 351 models, just behind Claude Haiku 5.5 at 43.4. Its Terminal-Bench 4.0 score there is 25.8%, lower than MiniMax reports.

The weakness is in the same data. On AA-Omniscience it gets 56.2% of answers right, ranked #9 of 280, but its non-hallucination rate is 7.2%, ranked #256 of 280. In plain terms: when it doesn't know, it almost never says so. The itplain reviewer saw exactly that. Asked about a city that does not exist, and told to say "I don't know" if unsure, it said the city doesn't exist and then invented a district with a population of about 189,000.

That fits a pattern we keep seeing behind viral AI claims, from the Kronos trading model's 93% figure to the Gemini 4 Argon quote: the headline is real, and the useful part is in the fine print. For M3.1 Flash, the fine print says use it for work you can check, like refactors with tests, UI builds you can look at, or data extraction you can diff, and keep facts and research on a model that admits uncertainty.

How do you switch from the free model to the paid one?

Change one string. Every route uses the same base model name once the free variant is gone:

# OpenRouter (was minimax/minimax-m3.1-flash:free)
minimax/minimax-m3.1-flash

# Vercel AI Gateway (free rate ended; same name now bills)
minimax/minimax-m3.1-flash

# Ollama (cloud only; now pay-as-you-go)
ollama launch claude --model minimax-m3.1-flash:cloud

# Self-hosted with vLLM 0.33.0+ on 8x H100/H200 (from the vLLM recipe, trimmed)
vllm serve MiniMaxAI/MiniMax-M3.1-Flash --tensor-parallel-size 8 \
  --block-size 128 --kv-cache-dtype fp8 --max-model-len 262144 \
  --reasoning-parser minimax_m3 --tool-call-parser minimax_m3 --enable-auto-tool-choice

MiniMax recommends temperature=1.0 and top_p=0.95. The vLLM recipe notes that --block-size 128 is mandatory because of the model's sparse attention. Before you switch a long-running agent to the paid slug, set a spending limit on the account. Always-on reasoning plus a 1M context window can turn a cheap per-token price into a surprising invoice.

Key facts

  • Model: MiniMax M3.1 Flash, released October 19, 2026. About 428B total parameters, 23B active, 1M context, text/image/video in, text out.
  • Free period: ended October 22, 2026 (OpenRouter 12:00 UTC, Ollama Cloud 23:59 UTC, Vercel AI Gateway Oct 22). Not in OpenCode's model list on Oct 23.
  • Price now: $0.15 input / $0.60 output / $0.015 cached per 1M tokens. Cheapest OpenRouter host Oct 23: inference.net $0.10 / $0.55.
  • Weights: Hugging Face, public Oct 23. 48 shards, 250.1 GB, plus a 2.3 GB DSpark draft head.
  • Hardware: vLLM recipe uses 4x B200/GB200/B300 or MI355X, or 8x H100/H200. Ollama has a cloud tag only.
  • License: MiniMax Community License. Free non-commercial; commercial needs "Built with MiniMax M3.1-Flash" plus a notice; over $20M revenue needs written approval.
  • Quality: Artificial Analysis Intelligence Index 42.7 (#21 of 351); non-hallucination rate 7.2% (#256 of 280).
  • X: launch post 1,276,369 views / 7,429 likes; weights post 108,052 views / 1,538 likes (checked Oct 23).

FAQ

Is MiniMax M3.1 Flash still free?

Not on the big routes. The free launch period ended on October 22, 2026: OpenRouter retired the :free model at 12:00 UTC, Vercel AI Gateway's offer ran until October 22, and Ollama Cloud's ended at 23:59 UTC. On October 23 the API costs $0.15 per million input tokens and $0.60 per million output tokens. The weights are free to download for non-commercial use, but running them needs a multi-GPU server.

When did the MiniMax M3.1 Flash free period end?

October 22, 2026. OpenRouter said its free period ended Thursday, October 22 at 12:00 UTC, 72 hours after launch. Ollama Cloud's free window ran until 23:59 UTC the same day, and Vercel's AI Gateway changelog said free until October 22. OpenCode's model list on October 23 no longer includes M3.1 Flash.

How much does MiniMax M3.1 Flash cost now?

MiniMax lists a flat $0.15 per million input tokens, $0.60 per million output tokens and $0.015 per million cached input tokens. On OpenRouter, inference.net was the cheapest host on October 23 at $0.10 input and $0.55 output. A job that reads 1 million tokens and writes 100,000 costs about $0.21 at MiniMax's rate.

Where can I still use MiniMax M3.1 Flash for free?

Three places, each with a catch. MiniMax Agent gives free accounts registration and check-in credits, which run out. AnyRouter, a small third-party router, still listed the model at $0 on October 23 with no end date, so read its data terms first. And the open weights are free for non-commercial use if you have the hardware.

Can I run MiniMax M3.1 Flash locally?

Not on a laptop or a single gaming GPU. The Hugging Face repo holds about 250 GB of weights, and the vLLM recipe serves it on 4 Blackwell-class GPUs (B200, GB200, B300) or AMD MI355X, or 8 H100 or H200 GPUs. Ollama only offers a cloud tag for this model, so ollama run sends your prompts to Ollama's servers.

Is MiniMax M3.1 Flash open source?

It is open weights, not open source in the OSI sense. The MiniMax Community License allows free non-commercial use. Commercial use requires a visible Built with MiniMax M3.1-Flash notice and a one-time email to MiniMax, and companies with more than $20 million in yearly revenue need written approval first. The license also lists prohibited uses, including any military purpose.

Is MiniMax M3.1 Flash good for coding?

On MiniMax's own table it scores 73.7% on DeepSWE v1.1, close to Opus 5 at 74.0%, but 38.1 on Terminal Bench 4.0 against 51.8 for Opus 5. Independent testing by Artificial Analysis puts it at 42.7 on its Intelligence Index, #21 of 351 models, with a 7.2% non-hallucination rate, which means it rarely says I don't know. Use it for cheap, checkable tasks and review the output.

Sources

Keep reading