MiniMax M3.1 Flash vs DeepSeek V4.1 Flash 1017 vs GLM-5.3-Flash: which Flash model is cheapest per finished task?

Last Updated: 23 October 2026

Key takeaways

MiniMax published the full weights for M3.1 Flash on 23 October 2026: 428B total parameters, 23B active per token, 1M context, native image and video input.

On MiniMax's own comparison table, the DeepSeek V4.1 Flash 1017 checkpoint beats M3.1 Flash on the coding benchmark MiniMax headlined, 77.0 against 73.7 on DeepSWE v1.1.

All three models sit within four points of each other on the Artificial Analysis Intelligence Index: MiniMax 43, GLM-5.3-Flash 42, DeepSeek V4.1 Flash 39 at max effort.

Cost per finished task is not cost per token. Artificial Analysis puts MiniMax at US$0.22, GLM-5.3-Flash at US$0.28 and DeepSeek at US$0.40 per Intelligence Index task, because DeepSeek emits 250M output tokens where the other two emit 180M.

Only MiniMax fits on a pair of DGX Sparks at its native release size (252.4 GB of weights). GLM-5.3-Flash needs 328.4 GB and DeepSeek V4.1 Flash needs 510.3 GB.

DeepSeek and GLM ship under MIT. MiniMax ships under a community licence that requires a "Built with MiniMax M3.1-Flash" credit, a notice email, and written authorisation above US$20M in yearly revenue.

MiniMax's launch tweet on 23 October pulled 256,384 views and 1,960 likes within eleven hours, and the numbers in it are real. What the tweet does not say is that two cheaper rivals already sit inside the same performance band, and that one of them wins the benchmark MiniMax chose to headline.

What did MiniMax actually release with M3.1 Flash?

MiniMax released the weights for M3.1 Flash on 23 October 2026, three days after the API launch on 19 October. According to the model card on Hugging Face, M3.1 Flash is a natively multimodal mixture-of-experts model with 428B total parameters and 23B activated per token, a 1M token context window, and text, image and video inputs. It carries MiniMax Sparse Attention through the whole model, quantization-aware training on the routed experts and attention paths, and a DSpark speculative decoding draft head.

The interesting part is not the size. It is the efficiency story. MiniMax's research post claims online inference is "nearly three times faster" than its predecessor and that usage cost dropped 70%, with output speeds reaching "200+ TPS" and a median of 170+ TPS. The flat price is US$0.15 per million input tokens, US$0.60 per million output tokens and US$0.015 per million cached tokens.

The model also shipped anonymously first. MiniMax says the stealth release, codenamed Space Bunny, processed more than 200 trillion tokens before anyone knew whose model it was. That detail matters later, because the people who used Space Bunny for free were not impressed.

How do MiniMax M3.1 Flash, DeepSeek V4.1 Flash 1017 and GLM-5.3-Flash compare on specs?

All three are Flash-class in name only. Each is a 300B-plus mixture-of-experts model with a 1M context window and native multimodal input, and each activates between 8B and 23B parameters per token. According to the respective model cards, MiniMax is the largest by total parameters, GLM-5.3-Flash is the smallest and the only one with a plain MIT licence and a hybrid sparse plus linear attention stack, and DeepSeek V4.1 Flash is the most aggressive on KV cache compression at 890 bytes per token.

Two details are easy to miss. First, DeepSeek's activation count is asymmetric: 8B during prefill and 16B during decode, which is why input-heavy agent workloads are cheap on it. Second, MiniMax's 252.4 GB release is smaller than GLM's 328.4 GB despite 108B more parameters, because quantization-aware training puts most of the expert weights in 8-bit storage at release time.

What do the benchmark tables say, and where should you distrust them?

MiniMax's table is unusually candid for a launch: it lists DeepSeek V4.1 Flash twice, once as the original release and once as the 1017 checkpoint. Read it closely and the headline claim weakens. On DeepSWE v1.1, the coding benchmark MiniMax quoted in its tweet, M3.1 Flash scores 73.7 while DeepSeek V4.1 Flash 1017 scores 77.0. On NL2Repo-Bench the gap is wider, 61.2 against 68.2, and on Terminal Bench 4.0 it is 38.1 against 42.3. MiniMax wins on Automation-Bench (62.8 against 60.5), on HLE without tools (43.4 against 42.6), and on the tool-heavy benchmarks DeepSeek did not report.

The CyberGym row deserves its own paragraph. MiniMax's post claims "leading capabilities in vulnerability discovery and validation" at 82.2, and its table shows a dash for both DeepSeek columns. DeepSeek's own model card reports CyberGym at 88.1 for V4.1 Flash and 84.5 for GLM-5.3, both above MiniMax's headline number. The dash is not proof of a worse score, but a reader comparing launch posts would conclude the opposite of what the vendor data shows.

Then there is the benchmark itself. DeepSWE v1.1 has been formally criticised: according to Epoch AI's review of the benchmark, 23 of 113 tasks had issues, just over 20%, and Epoch designated the benchmark as flawed. "I find it interesting that so many models seem to be at almost exactly 74% specifically which makes me suspicious of the benchmark," wrote Hacker News commenter penagwin on the launch thread. On the same thread, lambda pointed to whatmodel.sh's analysis that a lab's own benchmark runs come out about 6.4% better than independent reruns of the same evaluation.

The stealth release is the strongest counter-signal. "This model seems to be pretty benchmaxed," wrote DCKing on Hacker News. "In actual use, it doesn't survive the comparison to models in the benchmark." Several commenters said they used Space Bunny free on OpenCode Go and went back to DeepSeek V4 Flash or Qwen 3.8 Flash Next. The most balanced independent read came from Plotly's chriddyp, who benchmarked it on short-horizon data analytics questions: "it got the best score of all of the Chinese models and tied Opus 5.5 and Argon (41/43 questions correct)... it seems to use A LOT of tokens and it's the slowest model that we've tested for the exam."

One vendor claim does survive scrutiny. On real-world security work, eskibars, who runs a cybersecurity platform called ZeroQuarry, wrote that MiniMax "far underperforms relative to GLM 5.3 from every test I've done for real-world cybersecurity tasks", which lines up with GLM-5.3's higher CyberGym score.

Which model is cheapest per finished task?

Per-token price is a poor guide here, because the three models disagree on how many tokens a task needs. According to Artificial Analysis, measured at each model's maximum reasoning effort on 23 October 2026, MiniMax M3.1 Flash costs US$0.22 per Intelligence Index task, GLM-5.3-Flash costs US$0.28 and DeepSeek V4.1 Flash costs US$0.40. DeepSeek emits 250M output tokens across the index where the other two emit 180M, a 39% verbosity penalty that erases its cheaper cache rate.

Work the arithmetic on a single agent turn. Take 40,000 cached context tokens and 12,000 output tokens. On MiniMax that is US$0.0006 of cached input plus US$0.0072 of output, or US$0.0078. On GLM-5.3-Flash at list price it is US$0.0012 plus US$0.006, or US$0.0072. On DeepSeek at list price it is US$0.00024 plus US$0.0144, or US$0.0146, and if the same task needs 39% more output tokens as the Artificial Analysis data suggests, closer to US$0.020. DeepSeek's 98% cache discount is real, but output dominates an agentic workload and output is where DeepSeek charges double.

The marketplace changes that picture. OpenRouter lists the DeepSeek V4.1 Flash 1017 snapshot at US$0.15 input, US$0.60 output and US$0.003 cached, which matches MiniMax on input and output and beats it five times over on cache. If you buy through a reseller rather than DeepSeek's own API, the price gap disappears and DeepSeek's 209 tok/s becomes the tiebreaker. That is the pattern commenter 0xbadcafebee described on Hacker News: "DeepSeek's price is lower than GLM on input tokens, but higher than GLM on output tokens", so output-heavy reasoning work costs more on DeepSeek than on GLM.

Can you run any of them on your own hardware?

Two DGX Sparks give you 256 GB of unified memory between them, which turns the weights-on-disk figure into a buying decision. MiniMax M3.1 Flash is the only one of the three that fits a pair at its native release size, at 252.4 GB. GLM-5.3-Flash needs 328.4 GB, so three Sparks. DeepSeek V4.1 Flash needs 510.3 GB, so four Sparks at native size or a serious quantisation pass to fit two.

In our own dual DGX Spark rig we serve DeepSeek V4.1 Flash 1017 and measured 35 tokens per second on a single stream and 62 tokens per second at four concurrent requests in October 2026. That is a workable agent speed for a small team, roughly a third of the 209 tok/s Artificial Analysis records from DeepSeek's hosted API, and it costs nothing per token after the hardware is bought.

Be honest about the maths. Two DGX Sparks at about US$6,950 each do not beat a US$200 per month hosted plan on subscription arithmetic alone, and anyone who claims otherwise has not run the numbers. Local wins under four conditions: several staff sharing one agent, API volume high enough that per-token billing hurts, data that cannot leave your control, and work a hosted plan will not permit. For Australian firms there is a fifth: Privacy Act 1988 APP 8 requires reasonable steps before personal information goes to an overseas recipient, and a model running in your own rack removes that question entirely.

Which Flash model should you use for which job?

Treat the three as a portfolio, not a winner. MiniMax M3.1 Flash is the best fit for tool-heavy enterprise workflows and computer use, where its BankerToolBench 76.2, OSWorld 61.7 and Agents' Last Exam 29.6 are the strongest numbers in this group and the only ones reported. DeepSeek V4.1 Flash 1017 is the best fit for long-context coding and repository work, where its DeepSWE 77.0 and NL2Repo 68.2 lead, and its 209 tok/s matters most. GLM-5.3-Flash is the best fit where licence simplicity and output price dominate, at US$0.50 per million output tokens under MIT.

Match the model to the data, not the benchmark chart. Cloud-safe work, like public documentation, open-source refactors and marketing drafts, can go to whichever model is cheapest that week. Hybrid work, like client codebases with no personal information, can use a hosted API with zero data retention or a local model. Never-cloud work, like health records, legal files and financial statements, should stay on hardware you control, which today means MiniMax on a Spark pair or DeepSeek at a heavier quantisation.

What are the risks?

Three risks matter more than the benchmark spread. The first is licence drift: DeepSeek and GLM are MIT, MiniMax is not. The second is benchmark inflation, with a formal flaw finding against DeepSWE and a measured 6.4% vendor-to-independent gap. The third is verbosity, which quietly doubles your bill when a model reasons three times as long as its competitor to reach the same answer.

On the licence, MiniMax's terms are workable for most businesses. Commercial use requires the credit line "Built with MiniMax M3.1-Flash" on a website, interface, blog post or product documentation, plus a one-time notice email to [email protected]. Written authorisation is only required above US$20M in yearly revenue. For an Australian SMB that is one email and one footer line, not a legal project.

On benchmarks, the practical rule is to rerun your own. Pick twenty real tasks from your backlog, run them through all three models with the same harness, and count finished tasks and tokens burned rather than reading a leaderboard. Our experience matches the Hacker News sceptics: the model that wins a chart is not always the model that finishes the ticket.

Frequently asked questions

Is MiniMax M3.1 Flash better than DeepSeek V4.1 Flash?

It depends on the task. On MiniMax's own comparison table, DeepSeek V4.1 Flash 1017 scores higher on DeepSWE v1.1 (77.0 against 73.7), NL2Repo-Bench (68.2 against 61.2) and Terminal Bench 4.0 (42.3 against 38.1). MiniMax scores higher on Automation-Bench (62.8 against 60.5) and reports strong computer-use numbers that DeepSeek did not publish. On the independent Artificial Analysis Intelligence Index, MiniMax leads narrowly at 43 against DeepSeek's 39.

Can I run MiniMax M3.1 Flash on two DGX Sparks?

Yes, at the native release size. The Hugging Face repository holds 252.4 GB of weights across 48 shards, and two DGX Sparks provide 256 GB of unified memory. It is a tight fit with little headroom for long context, so plan on a quantised build if you need the full 1M token window.

What does the MiniMax community licence allow for commercial use?

Commercial use is permitted with two conditions: display "Built with MiniMax M3.1-Flash" on a related website, interface, blog post or product documentation, and send a one-time notice to [email protected]. Above US$20M in yearly revenue you need prior written authorisation from MiniMax. Non-commercial use is free of both conditions.

Which is cheaper for coding agents, GLM-5.3-Flash or DeepSeek V4.1 Flash 1017?

It depends on where you buy and how verbose the task is. GLM-5.3-Flash lists at US$0.50 per million output tokens against DeepSeek's US$1.20 list price, so GLM is cheaper on output-heavy coding. Artificial Analysis measures US$0.28 per task for GLM and US$0.40 for DeepSeek at max effort. Through OpenRouter, the DeepSeek 1017 snapshot is listed at US$0.60 output and US$0.003 cached input, which closes most of the gap.

Do these benchmark scores hold up in real work?

Partly. DeepSWE v1.1, which carries most of the headline claims, was reviewed by Epoch AI with issues found in 23 of 113 tasks and designated as flawed. Users who tested MiniMax M3.1 Flash anonymously as Space Bunny reported mixed results on real coding work, while Plotly's independent data-analytics benchmark placed it first among Chinese models at 41 of 43 questions. Rerun your own tasks before committing a team to any of the three.