Two cheap models you can actually run: MiniMax M3.1 Flash and DeepSeek V4.1 Flash

Three months ago the argument was that cheap models were not good enough to run real work, and expensive models were good enough to do everything. That gap has narrowed. Not to nothing, but enough that the question for a small business is no longer "which is the best model" but "which one can I afford to leave running, and can I run it myself".

These two releases are the clearest evidence yet, and they are worth understanding separately because they answer different halves of that question.

1. MiniMax M3.1 Flash: the preview finally became a real model

MiniMax shipped an M3.1 Flash preview on 27 September with no benchmarks attached, which made it hard to plan around. The general release on the 19th of October is the one with numbers behind it, and those numbers are good.

It is a mixture-of-experts model with 428 billion total parameters and about 23 billion active, a one million token context window and native multimodal input. Output speeds reach 200 tokens per second, with the median consistently above 170. MiniMax says agentic task costs are 70 percent lower than the preview it replaces, and that online inference is nearly three times faster.

The provenance detail is worth a line: Space Bunny, the agent that quietly processed more than 200 trillion tokens while running anonymously, was an early version of this model. That is a claim about sustained production use, not a benchmark.

2. What MiniMax's own benchmark table says

These are the vendor's numbers, published on the 19th. The useful reading is not the absolute score but where it lands against the frontier models.

• BankerToolBench 76.2, tied with Opus 5. This is the multi-step, document-heavy, compliance-flavored work, and tying the expensive model here is the headline result.
• DeepSWE v1.1 at 73.7 percent, just above GPT-5.6 Sol at 73.0, and behind DeepSeek V4.1 Flash 1017 at 77.0.
• OSWorld V2 at 61.7 and Agents' Last Exam at 29.6, first among Chinese models on both, but behind Opus 5 at 70.6 on OSWorld.
• CyberGym at 82 percent, with the vendor reporting strong results on code auditing, vulnerability patching, penetration testing and malware analysis.
• Automation-Bench 62.2 and LOCA-Bench 60.3, both ahead of Opus 5 at 50.3 and 53.7. This is enterprise workflow execution, which is closer to what a small business actually buys.

And the weak spot, stated plainly: Terminal Bench 4.0 is 38.1 and SWE Refactor Bench is 23.0 against Opus 5 at 31.0. Long autonomous terminal sessions and repository-scale refactoring are where it trails. That is a real limit, not a rounding error, and it is the class of work most likely to matter if you are asking an agent to work unattended for hours.

3. DeepSeek V4.1 Flash 1017: a small update with a big licence

DeepSeek upgraded V4.1 Flash to a checkpoint numbered 1017 on the 17th of October. Their own documentation calls it a minor update. The benchmark table on the same page says otherwise.

• Terminal-Bench 4.0 moved from 31.2 to 42.3, a jump of over eleven points on the hardest agentic benchmark.
• DeepSWE v1.1 from 74.2 to 77.0, which is higher than both MiniMax M3.1 Flash and GPT-5.6 Sol.
• Apex Bench v1.1 from 65.6 to 72.1, and Agents' Last Exam from 31.8 to 35.4.
• Automation-Bench from 54.8 to 60.5, and ProgramBench from 20.3 to 29.1.

DeepSeek also reduced the model's task runtime and output length, so it does the same work for less money and less waiting. They describe it as a new performance and cost Pareto point, which is the phrase that should worry anyone charging per token for a living.

The context window is one million tokens with up to 384 thousand tokens of output, and it handles images natively. Requests using the model name deepseek-flash are routed to the 1017 checkpoint automatically, so nothing needs changing in your code to get it.

4. The part that matters most: the weights are MIT licensed

DeepSeek published the weights for V4.1 Flash 1017 on Hugging Face under the MIT licence, with community GGUF and MLX builds already available. No licence fee, no per-seat charge, no vendor deciding in eighteen months what it now costs you.

The model is a 763 billion parameter mixture-of-experts design, so this is not a laptop project. The quantised builds are what make it practical on a workstation, and there are several. But if data leaving your building is a problem you have to solve rather than a preference you have, this is the first genuinely capable model where the answer is a download rather than a negotiation.

MiniMax is the other way round, and this is the detail to hold onto. The M3.1 Flash API is live now, but the weights are not published and the licence has not been disclosed. MiniMax says weights arrive on 23 October, and its own documentation already describes the model as open-weight. So expect a licence, a repository and a great deal of local-running interest within days. What you cannot do today is plan a self-hosted deployment around it.

5. The pricing, including the bit that is easy to miss

MiniMax M3.1 Flash is a flat $0.15 per million input tokens, $0.60 per million output and $0.015 per million cached input, and the flat rate applies whatever the length of the prompt.

DeepSeek looks identical until you read the price table properly. Off-peak it is $0.15 per million input, $0.60 per million output and $0.003 for a cache hit. At peak it doubles to $0.30 and $1.20. Peak hours are 01:00 to 04:00 and 06:00 to 10:00 UTC on weekdays, so in the UK that is the small hours of the morning and the start of the working day. Every other hour, including all of every weekend, is off-peak.

That is a free halving for anyone running overnight batch work, scheduled quotes or backfills, and it is the sort of detail that quietly halves a monthly bill for a small business doing a lot of scheduled AI work.

6. What this actually means if you are running AI in a business

The honest summary is that the middle of the market has got much better, and the top of the market has got further away in price but not in capability. That combination suits a particular shape of business.

It suits companies with high volume, repetitive, measurable work. Enquiry handling, quoting from a price list, support answers drawn from your documentation, CRM updates, reporting. Work where a mistake is a fix rather than a crisis, and where the same few thousand tokens repeated ten thousand times is where the arithmetic works.

It does not suit work that needs long unattended terminal sessions, because that is precisely where M3.1 Flash scores 38.1 against 42.3 and where both models trail the frontier. If your plan is an agent left alone overnight on a complex repository, you are still in expensive-model territory.

And there is one more change worth noticing. The gap between "frontier model" and "cheap model" used to be measured in capability. It is now increasingly measured in licence and availability, which is a much better problem to have. Two of the strongest cheap models released this month are one with a permissive open licence and no per-seat fee, and one that will be, imminently.

7. What to do this week

If you already run a cheap model in production, this is a good fortnight to re-benchmark it. The AI automation ROI calculator will tell you what a price change is worth in your specific case, because the answer depends entirely on your token volume and cache hit rate.

If you have been waiting to run something on your own hardware rather than sending it to an API, DeepSeek V4.1 Flash 1017 is the first one where the licence makes that a decision rather than a gamble. The set up your own AI agent guide walks through the setup, and your data, your rules covers what you can and cannot safely put into a hosted model.

If you are running Hermes itself, the v0.22.0 release added one-click local models, so a quantised DeepSeek or Qwen on your own hardware is now a menu item rather than a weekend. That is covered in what changed in Hermes 0.22.0.

If you want to know whether any of this is worth doing at all for your specific business before spending a weekend on it, the where AI actually helps guide is the place to start.

8. The honest take

Neither of these is the model you should build a strategy around. Both are models you should have costed, because the price of the boring work just fell again and the gap between the cheap end and the expensive end is now measured in capability points rather than multiples of your bill.

The uncomfortable part for anyone selling AI by the seat is that MIT-licensed weights with no per-user charge make that business model harder to defend every time something like this ships. The model you run becomes a decision about your hardware and your data, not a line on a corporate card.

That is good for the buyer. It is also why the July piece on Chinese labs shipping faster and cheaper is worth re-reading now, because the pattern in it held, and it held again this month.