ignis is deliberately narrow: one model family, one class of card. It
serves Qwen3.8 (the 27B and the 125B Flash-Next MoE) and Qwen-Image-2.1 on a
single Blackwell card, an RTX 5090 or an RTX PRO 6000, and it gives up
generality to take back speed. Everything in it is shaped around the load a
developer actually produces: one main agent plus a handful of subagents,
hitting the same card at once, with long contexts, tool calls and reasoning.
These pages tell what that buys, one topic at a time. Each one says how it
works, what was measured and on what, how you see it and switch it, and where
it stops. Every number links to the finding that measured it.
Measured on one RTX 5090 with 64 GB of DDR4-3200 on a PCIe 3.0 board, the
27B NVFP4 unless a tile says otherwise. How these numbers are made
says what each kind of figure is, and what it is not.
One round for every laneEight lanes share one pass over the weights, replayed from a CUDA graph per batch width, at the card's memory bandwidth.
A drafter inside the artifactDFlash2 proposes up to seven tokens a round and the model verifies them in the same pass. The output stays the model's own.
A 125B MoE on a 32 GB cardFlash-Next's experts in a 2.5-bit trellis code, the hot ones in VRAM and the rest streamed from pinned RAM.
The 27B on a 16 GB cardA 3-bit trellis body, a profile chosen by the card's memory, and four quality gates passed against the NVFP4.
KV at 9 KB a tokenA compressed KV format holds 7.11x the tokens of BF16, budgeted in bytes. Its quality cost is measured, not hidden.
State that outlives its requestA repeated prompt resumes from a copy held in VRAM, pinned RAM or on disk, and the response says how much it reused.
Every byte planned at loadThe whole VRAM plan is reserved and printed before the first request. A start it cannot hold is refused, not failed at minute forty.
Lanes that know who is askingA request says whether it is the conversation or a subagent. A burst of subagents fills idle lanes and yields first when memory is short.
Reasoning that finishesA loop is caught as it streams, cut back and nudged in the same response; a turn that ends inside its thinking is closed, not returned empty.
System One decisions, without decodingThe System One API (POST /v1/systemone) on a local model: a yes/no, a choice or a score read from one position of the model's distribution, one prefill, nothing generated.
Images as evidenceThe vision tower's output outlives the request: four questions over one screenshot encode it once.
