Pyroeis 3 is the base route in the Sigilix model line. It is built for the work you ask for all day: explain the changed code, summarize the issue, find the likely failing path, draft the small repair, and keep the answer grounded in the repository.
Pyroeis is shaped by the product layer that matters in practice: codebase memory, workflow context, and the patterns your workspace has already taught the system to recognize.
Base route benchmark frame
Pyroeis 3 is shown as the Sigilix base route. The comparison keeps the reference base-tier peers in view while separating task scores, output capacity, and context capacity.
These charts are capability context, not a raw-model leaderboard or an isolated Pyroeis-only benchmark run. Pyroeis is designed for everyday work where your repository memory, issue history, and workflow context make the route more useful in the actual product.
TerminalBench measures command-line coding work: reading the repository, selecting a next action, responding to shell feedback, and continuing through a task without losing the state of the workflow.
That is the everyday surface Pyroeis is built for. You are often not asking for a perfect research answer; you are asking what failed, what changed, where the bug likely sits, or what command moves the task forward.
The Pyroeis row starts from the reference TerminalBench 2.1 base value, then displays the route with the 3.6-point Sigilix context lift applied. Some peers only publish TerminalBench 2.0-style rows, so those rows are labeled with the available version rather than being treated as identical evidence.
The practical read is not that Pyroeis outranks every base peer. The practical read is that the base lane is strong enough to stay useful while the product gives it the local memory and repository context that raw benchmark prompts do not include.
Values shown: GPT-5.6 Terra TerminalBench 2.1; Base peer A TerminalBench 2.1; Pyroeis 3 TerminalBench 2.1; Base peer B TerminalBench 2.1; Qwen 3.7 Plus TerminalBench 2.0; Qwen3-Coder-Next TerminalBench 2.0; Nemotron 3 Super TerminalBench 2.0.SWE-bench Verified is closer to repository repair: find the relevant code, understand the failure, make the patch, and preserve the constraints that tests are trying to enforce.
For Pyroeis, that matters because base-model work still has to be believable. A quick answer is not useful if it points at the wrong file, forgets a product constraint, or ignores the convention the repository already established.
The Pyroeis row is shown at 85.0% on SWE-bench Verified with the Sigilix context lift already reflected in the route value. Sigilix does not present this as a detached model win. The product value is that the route can reuse your local codebase memory so each repair question starts closer to the actual work.
That is why the base lane is important. Most day-to-day repairs are not novel research problems; they are small, repeated patterns that become safer when the tool remembers what your codebase already taught it.
Values shown use the reference SWE-bench Verified rows for Base peer A, Base peer B, Pyroeis 3, Qwen 3.7 Plus, Qwen3-Coder-Next, and Nemotron 3 Super.Output capacity is not accuracy. It tells you how much room the route can use when it needs to explain a change, propose a patch, or carry a repair plan through multiple steps.
Pyroeis does not need the largest output cap to be useful. The base route is meant to answer the everyday question cleanly, then let the surrounding product memory carry the context that would otherwise be repeated in the prompt.
The chart uses the documented max output-token limits for each row. The effort chart that follows separates raw output budget from effective repeated-context load, because Pyroeis applies the tested 63% reduction at each effort point.
Values shown use documented max output-token limits from the capacity references for Base peer B, Base peer A, Pyroeis 3, Qwen 3.7 Plus, Qwen3-Coder-Next, and Nemotron 3 Super.This view shows the same base-route comparison as an effort curve. Peer lines step through public output-budget checkpoints; Pyroeis plots the effective repeated-context load after the tested 63% reduction at every effort point.
For Pyroeis, the important part is the shape of the route. Minimal effort is for quick answers, low and medium effort add room for evidence and patch planning, and high effort reaches the context-lifted route score while carrying less repeated context through the prompt.
The peer lines are shown as capacity checkpoints rather than independent per-effort vendor claims. That keeps the chart honest: the reference benchmark values stay at the row level, while the visual explains how output budget changes the space available for reasoning.
Only the rightmost point on each line is the benchmark anchor. Earlier points show the route's working shape as output budget increases, not separate published benchmark claims.
Hover a Pyroeis dot to see both the effective output tokens plotted on the chart and the raw output budget before the 63% repeated-context reduction.
Scores are anchored to the reference SWE-bench Verified rows. Pyroeis effective tokens use the tested 63% repeated-context reduction; peer output-token and context-window limits come from the capacity references for Base peer A, Base peer B, Qwen 3.7 Plus, Qwen3-Coder-Next, and Nemotron 3 Super.Context capacity is one reason Pyroeis works as a base route. A large context window gives the model room to inspect more of the active surface before the product memory layer has to decide what to compress, retrieve, or carry forward.
The important distinction is that context capacity and memory are not the same thing. Context is what can fit in the current run; memory is what Sigilix can preserve across runs so you do not have to keep rebuilding the same background.
Pyroeis is designed for the middle of that workflow: enough native context to answer everyday questions, plus enough memory support to make repeated work improve inside your org or user scope.
Values shown use documented context-window limits for Pyroeis 3, Base peer A, Qwen 3.7 Plus, Nemotron 3 Super, Base peer B, and Qwen3-Coder-Next.Pyroeis 3 is the base lane for everyday engineering work. The benchmark rows come from the TerminalBench, SWE-bench Verified, max-output, and context-capacity references for the route and peer rows, with the displayed Pyroeis task scores carrying the Sigilix context lift. The effort graph shows effective output load after the tested 63% repeated-context reduction at each Pyroeis effort point. The product claim is about how Sigilix uses local memory and codebase context around that route, not about training a shared model on customer data by default.
everyday engineering work
with context lift
with context lift
native window
What Pyroeis is for.
Pyroeis is the everyday lane. It is for small and medium coding questions where you need a useful answer quickly, but still need the answer to understand the repository instead of giving generic advice.
That includes review summaries, test-failure explanations, Linear issue breakdowns, CLI follow-ups, local code-path questions, and repair plans that should preserve the conventions already present in your codebase.
The route is not meant to turn every request into a frontier-model event. Most engineering work is not a once-a-quarter benchmark task; it is a stream of small decisions that get easier when the tool remembers what happened five minutes ago, yesterday, and last sprint.
Pyroeis is tuned for that surface. It should be quick enough to stay in your flow, deep enough to carry evidence, and grounded enough that the answer can point back to files, decisions, and constraints instead of floating above the work.
How memory changes the route.
Pyroeis is useful because the model is not asked to start from a blank prompt every time. Slack decisions, Linear issue history, prior CLI sessions, review outcomes, and repository conventions can all become context for the next answer.
That does not make the public benchmark row disappear. It changes the product behavior around the model. The system can spend less time rediscovering the shape of the repository and more time applying the right pattern to the current task.
This is the training and routing direction behind the product: models should learn common codebase failures and then reason with the evidence system around them. The memory layer is how those lessons become useful inside your workspace.
The improvement is scoped to the org or user context available to the task. Sigilix does not train product-wide models on customer data unless that is enabled; the default product value comes from carrying your local context forward.
Benchmark framing.
The benchmark section keeps four questions separate: terminal workflow performance, repository repair performance, max output capacity, and context capacity. Those are related, but they are not the same claim.
TerminalBench is useful for command-driven work. SWE-bench Verified is useful for repair work. Output capacity tells you how much the model can emit when it needs to explain or patch. Context capacity tells you how much surface the route can hold before the product memory layer starts doing more of the work.
Some peer rows use TerminalBench 2.1 and some use TerminalBench 2.0 because that is what is available in the reference dataset. The chart labels that difference instead of smoothing it away, and the rows should not be read as a normalized benchmark leaderboard.
Read the table as a capability frame for an everyday route. Pyroeis is not presented as a raw benchmark leader or as an independent Pyroeis-only benchmark run; it is the base route for routine engineering work where long context and workspace memory can make the answer feel less disposable.
Why the base lane matters.
A strong base route changes how often you can ask for help. If every question has to wait for the heaviest route, people stop asking the small questions that prevent messy follow-up work.
Pyroeis is meant to sit closer to the work: enough context for real repository answers, enough speed for everyday use, and enough continuity that repeated questions do not feel like starting over.
When a task becomes broad, security-sensitive, or high-stakes, you can route deeper. The base lane is for the majority of daily work that needs believable context without making the interaction feel expensive.
What comes next
Pyroeis is the base route for the everyday work you should not have to explain twice.