MiniMax M3.1 Abliterated: 129.9 Tokens a Second at 0.272% Errors

MiniMax M3.1 Abliterated: 129.9 Tokens a Second at 0.272% Errors

Few cards on OrcaRouter‘s cyber tier combine serving rows this good with a footprint this small. MiniMax M3.1 Flash abliterated runs at 129.9 tokens a second, 0.272% errors and a 1,695 ms median, on a daily latency series that barely moves [OURS].

It also carries 2.74 billion tokens a week — a quarter of Kimi’s traffic and a twelfth of DeepSeek’s — while charging the same $0.80 in and $1.60 out as the leader.

That gap between performance and adoption is what this piece is about.

The seven daily medians

1,698, 1,715, 1,696, 1,729, 1,588, 1,727 and 1,646 ms across the seven days to 2026-11-15 [OURS].

A spread of 141 ms on a median of 1,695 — an 8% band. No other card in the tier comes close. GLM 5.3 Flash’s daily p50 runs from 3,477 to 4,589. MiMo’s runs from 4,312 to 10,000. DeepSeek’s runs from 895 to 1,416 on a much lower base [OURS].

A flat series means the card is provisioned for its load rather than absorbing peaks. For a capacity planner that is the most valuable single row in the tier, because it is the only one that predicts next week.

Why the flatness is worth more than the speed

A card at 129.9 tokens a second that varies by 8% is a card you can build a timeout around.

Set a 20-second client timeout and, with a 1,695 ms median and a 6,429 ms p95, almost nothing should hit it. The daily series says the same thing seven times: no day was materially worse than another.

Compare a card with a better headline number but a volatile series. DeepSeek’s median is 998 ms — 41% faster — but its daily series runs from 895 to 1,416, and its p95 is 8,938 ms [OURS]. A timeout sized for the median fails on the tail.

The choice between them is not speed. It is whether you want the fastest card on average or the most predictable one.

What the card gives up

Two rows, both structural.

Context window: 262,144 tokens. That is the smallest in the tier alongside Qwen3.8 Flash Next, and a quarter of what nine other cards offer [OURS].

Output ceiling: 131,072. Also a quarter of its base checkpoint’s 524,288 [OURS]. MiniMax is the only card in the tier that is smaller than its base on both axes.

So the card is a quarter of the checkpoint in two dimensions and the best in the tier in three others. It is a deliberate trade, and it is the clearest example on the shelf of an abliterated build that changed shape rather than only behaviour.

Where the traffic goes instead

DeepSeek carries 32.99 billion tokens a week against MiniMax’s 2.74 billion [OURS]. Twelve times the volume on a card that is slower in the median and errors more often in the tail.

The likely explanation is the window. A pipeline that sends a repository or a large report needs a million tokens, and MiniMax cannot take it. The card’s excellent serving rows are unavailable to any workflow that needs the context, which excludes most repository-scale security work.

That is a clean and unglamorous reason for a performance leader to be a volume laggard: it is not that the card is worse, it is that most jobs in this tier are larger than it can hold.

The jobs that fit

Three workloads where 262,144 tokens is not a constraint and the serving rows decide.

Triage. A finding, a log excerpt, a short report. Hundreds to tens of thousands of tokens in, a classification out. The 262,144 window is enormous for this, and the 0.272% error rate means the pipeline does not need retry logic.

Batch extraction. Indicators of compromise from a set of short documents. Fast, cheap, reliable, and the window barely matters.

Interactive review. A human asks questions about a single file and expects an answer in seconds. The flat 1.7-second median is the whole experience here, and the 8% band means it stays that way.

What does not fit: whole-repository review, long video analysis, and any job whose output exceeds 131,072 tokens.

The cost of one triage call, worked through

A triage step on this card is small and frequent, so the arithmetic that matters is per call rather than per month.

Take a 4,000-token log excerpt and a 200-token classification. At $0.80 in and $1.60 out that is $0.0032 in and $0.00032 out, so $0.0035 per call [OURS, our arithmetic on the published rates].

Ten thousand calls a day is $35.00. At the 0.272% error rate, roughly 27 of those calls fail, and at one retry each the extra cost is nine cents. Retry logic is not worth writing for this card.

Now the comparison that matters: the same 10,000 calls with the excerpt repeated as a stable prefix. If 3,000 of those 4,000 tokens are identical across calls — a schema, a template, a standing instruction — caching at $0.15 per million drops the per-call input cost from $0.0032 to $0.0013, and the daily bill from $35.20 to $15.70 [OURS].

That is the whole case for this card in two numbers. Its serving rows make the call cheap and its flat latency makes the queue predictable, and the cache row decides whether the bill is $16 or $35 a day.

The benchmark rows, with their caveats

MiniMax publishes four: AA-LCR 79.7, GPQA Diamond 88.6, TAU-Bench Airline 68.0 and Vibe Code Bench v1.1 73.9, all evaluated 2026-10-18 — ten days before the card shipped [VENDOR].

That is the shortest gap between evaluation and release in the tier, which makes these rows marginally more current than the others. They still describe the checkpoint.

TAU-Bench Airline at 68.0 is the weakest of the four and the most relevant to agent work. A card that scores 88.6 on reasoning and 68.0 on tool use is a card you can trust with analysis and should test before trusting with a multi-step agent.

What a pilot should measure

Three numbers, all from your own traffic.

The p50 on your prompts. MiniMax’s 1,695 ms is measured on the harness’s prompt mix, which is not yours. A 50,000-token prompt will be slower.

The share of your jobs that exceed 262,144 tokens. If it is under 5%, the small window costs you almost nothing and the serving rows are free.

The error rate on your key. The published 0.272% includes the access gate; your provisioned key should see less.

A week of those three at $0.80 input costs less than the analysis that would replace them.

The short version

The best serving rows in the tier — 129.9 tokens a second, 0.272% errors, a flat 8% latency band — on a card that is a quarter of its own checkpoint in context and ceiling.

Use it for triage, extraction and interactive review, where 262,144 tokens is not a constraint. Do not use it for whole-repository work or long reports.

Its low traffic is explained by the window rather than by the performance. That is the single most useful thing to know before writing it off as underused.

Sourcing note: prices, context windows, output ceilings, modalities, quality scores, latency, throughput, error rates, traffic, ZDR flag and parameter surface are read from the card’s public model data on 2026-11-15, with seven-day windows ending that day [OURS]. Benchmark rows are OrcaCyber’s own, evaluated 2026-10-18 and inherited from the base checkpoint [VENDOR].

Similar Posts

Leave a Reply

Your email address will not be published. Required fields are marked *