OrcaCyber MiMo V2.6 Flash: Same Price as Its Sibling, Seven Points Better on CyberGym and Fifteen Worse on Exploit Reproduction

OrcaCyber MiMo V2.6 Flash Abliterated went live today on the same rate card as OrcaCyber Flash 1.0 — $0.80 in and $1.60 out per million tokens, a $0.15 cache read, a million-token context, refusal behaviour removed. Same price, same window, same absence of hedging. What separates them is where each one is strong, and on our own measurements the two profiles are close to mirror images: the new build leads CyberGym by seven points and trails its sibling by fifteen on exploit reproduction. Anyone weighing a model that engages directly with offensive-security work should also keep the wider record in view, which is what the OrcaRouter AI Incident Archive exists to collect. This post sets out what shipped, what the numbers are, and which workloads each build actually earns.

What actually shipped

MiMo V2.6 Flash is Xiaomi's efficiency checkpoint from its September 2026 MiMo-V2.6 release: a sparse mixture-of-experts model with 309 billion parameters in total and 15 billion active per token, MIT-licensed, with a million-token context. The OrcaCyber build is that checkpoint with refusal behaviour edited out of the weights — abliteration, not fine-tuning: no retraining corpus, no new architecture, just the directions the model refuses along subtracted from the matrices that carry them. It is the third name in the OrcaCyber line, after Nano and Zero, and it is the one built for throughput.

The input surface is wider than its sibling's. MiMo V2.6 Flash accepts text, image, audio and video and returns text, so a screen recording of a console session, a captured call, a network diagram or a screenshot of an error state can all be reasoned over in the same request as source code and a long log. Output is capped at 131,072 tokens per response. Thinking and non-thinking modes are both available with selectable reasoning effort, alongside native function calling and structured JSON output, and the model is reachable over Chat Completions, OpenAI Responses and Anthropic Messages.

The price is identical, so the price decides nothing

Both models sit at $0.80 input and $1.60 output per million tokens, with cache reads at $0.15 — the same card, to the cent, and no markup added on our side. That makes this an unusually clean choice. There is no cheaper model to talk yourself into and no premium tier to justify. The only question is which set of rows matches the work in front of you.

Access is the other thing the two share, and it is not open to any caller: both routes sit behind an engagement, a passkey and the terms, which is the right shape for a model whose purpose is authorized offensive work. Price and capability are public; the ability to call either one is not.

Two different bases under identical cards

The cards match because the price does not depend on the base, and the bases are not the same shape. MiMo V2.6 Flash is a 309-billion-parameter sparse model with 15 billion active per token; the DeepSeek checkpoint behind Flash 1.0 carries 552 billion parameters with 16 billion active. Both are MIT-licensed, both were edited at the weight level rather than retrained, and both arrive with refusal behaviour removed.

The pattern in the numbers is consistent with that difference without proving it: the smaller base starts faster and caps output lower, while the larger one does better on the rows that require carrying a long exploit chain to a working result. A model's parameter count is a poor predictor of any single benchmark, so treat this as context for reading the rows rather than as an explanation of them.

Where the new tier wins

• CyberGym — 95.1% for MiMo V2.6 Flash against 88.1% for Flash 1.0, a seven-point lead on the cyber-reasoning board.

• Median first token — 928 ms against 1,970 ms on our playground traffic, roughly twice as fast to start, with an output rate of 765.9 tokens per second against 171.1.

• Input types — audio and video on top of text and image, which the sibling does not accept at all.

The speed figure deserves a caveat before it is used to make a decision: it comes from one day of playground calls carrying about 17,500 tokens in total, while the sibling's numbers come from a seven-day window carrying 17.5 billion. A single day on a new route measures an empty queue more than it measures a model. Treat the direction — this build is the faster one — as reliable, and the exact multiple as provisional until the window fills.

Where it gives up ground

• SEC-Bench Pro — 47.5% against 62.8%, a 15.3-point deficit and the largest gap between the two models on any row we publish.

• DeepSWE v1.1 — 67.9% against 74.2%, 6.3 points behind.

• Terminal-Bench 2.1 — 87.6% against 90.6%, 3.0 points behind.

• Maximum output — 131,072 tokens against 393,216, a third of the room for a single long artefact such as a full audit report or an annotated exploit chain.

• Rows not yet published — ExploitGym (15.3%) and GPQA Diamond (90.9%) exist for Flash 1.0 only; the new build has no comparable figure yet.

The shape of that list is the argument. CyberGym rewards reasoning about an attack surface; SEC-Bench Pro rewards reproducing a known vulnerability end to end, and that is where the new model is materially weaker. If the work is triage, review, detection engineering or explaining what a log means, the fast tier is the better instrument. If the work is turning a CVE into a working proof of concept, the sibling's fifteen-point advantage is the number that matters, and no amount of latency makes up for a reproduction that does not land.

What these numbers are, and what they are not

Every benchmark in this post is ours. The model cards list CyberGym, SEC-Bench Pro, DeepSWE, Terminal-Bench, ExploitGym and GPQA Diamond with our own source field, and neither OrcaCyber model has been scored by an independent evaluator such as Artificial Analysis. That matters more for a security model than for a general one, because the benchmarks themselves are young and the gap between "reasoned about the vulnerability" and "reproduced the vulnerability" is exactly where a self-reported row can flatter a model. Read the numbers as vendor figures with a published methodology, not as audited results, and re-test on your own corpus before a production path depends on them.

How to choose

• Exploit development, proof-of-concept construction, patch validation — Flash 1.0. It is 15.3 points better on the reproduction row and 6.3 better on DeepSWE, and neither is a margin that evaporates under retry.

• Log and packet-capture triage, secure code review at volume, detection engineering, agent loops that need a first token fast — MiMo V2.6 Flash. The CyberGym lead and the latency class both point the same way.

• Long single-pass deliverables — Flash 1.0, purely on output ceiling: 393,216 tokens against 131,072.

• Screen recordings, call audio, diagrams and dashboards — MiMo V2.6 Flash, which is the only one of the two that accepts them.

Once access is enabled, both models sit behind one key, so the choice does not have to be permanent and does not require a second contract or a code change: route by task type, and let automatic failover cover the case where one route degrades mid-run. The workload rules above are ours; the figures they rest on are on the model pages.

What to watch

Three things will change this picture. The first is an independent score: until someone outside the catalogue runs these weights through a public harness, the fifteen-point exploit gap is our measurement and should be quoted as such. The second is the latency window; a week of real traffic on the new route will either confirm the two-times speed difference or shrink it. The third is the benchmark sheet itself — a model shipped today with four rows published and two rows missing is a model whose profile is still being drawn. The model page is the place that gets updated first.

Sourcing note: benchmark scores, prices, context and output limits are OrcaRouter's own catalogue entries for OrcaCyber MiMo V2.6 Flash Abliterated and OrcaCyber Flash 1.0, read on 26 October 2026, and are self-reported rather than independently audited; latency figures are our playground traffic — one day and roughly 17,500 tokens for the new build, seven days and 17.5 billion tokens for its sibling — and are not production benchmarks. Base-model details (309 billion total parameters, 15 billion active, one-million-token context, MIT weights) come from Xiaomi's MiMo-V2.6 release; the 552 billion total and 16 billion active parameters for the DeepSeek V4.1 Flash checkpoint behind Flash 1.0 are Artificial Analysis's record of that model. The workload recommendations and the interpretation of the benchmark rows are ours.