Glossary

Abliterated variant. qwen3.8-27b-abliterated, the 27B with a published abliteration applied: the same container with 70 matrices re-encoded, it does not refuse requests. See models and artifacts and the uncensored variant.

Acceptance. How many of a round’s drafted tokens the model agrees with, and so how many tokens a speculative round yields. See speculative decoding.

Admission. The step where the server decides whether a request runs, on which lane and in what rank, checking the KV pool and looking for retained state. See how ignis serves a request.

Agent. One of the two lane tags. An agent request, such as a subagent’s, ranks below an interactive one and is the first to move to host RAM when memory is short. See Lane tag.

Artifact. The .ninfer container a model is served from: weights, tokenizer and chat template in one file, checked against its sidecar. See models and artifacts.

Body format. How an artifact stores its transformer layers (NVFP4 or trellis). It is read from the artifact, never configured. See models and artifacts.

Alias. A second id a catalog entry answers to, within its own model family and only on a card it loads on: on a 16 GB card, qwen3.8-27b-16g answers to qwen3.8-27b. See Get a model.

Card. The GPU ignis runs on. The built-in profiles are written for the 16 GB class, a 32 GB RTX 5090 and a 96 GB RTX PRO 6000, and --profile auto picks one by the card’s memory. See profiles.

Card class. The cards one built-in profile serves, named by their memory (16, 32 or 96 GB), and the smallest one a catalog entry loads on. An entry above the card is refused before anything is unloaded or fetched. See models and artifacts.

Catalog. The models the server knows how to fetch: five built into the binary, pinned file by file, plus any an operator adds. See models and artifacts and Get a model.

Checkpoint. See Prompt checkpoint.

Chunk. The slice of a prompt that one prefill call processes (--model-prefill-chunk). Decode rounds run between chunks. See how ignis serves a request.

Committed run. The tokens a verify round emits for one lane: the anchor plus the accepted drafts, between 1 and the draft count plus one. See speculative decoding.

Continuation. The request a loop-guard rewind submits: the same input, extended by the reasoning kept before the cut and a nudge. See reasoning and the loop guard.

Deadline. How long a request may run before the server cuts it (--server-request-timeout, --server-stream-timeout). See deadlines.

Decision. An answer read from the logits at one position after prefilling the evidence, with no token decoded. It is the POST /v1/decide endpoint. See decisions.

Decode lane. See Lane.

Decode round. One traversal of the model that advances every decoding lane by a token, replayed from a CUDA graph for the exact number of lanes. See how ignis serves a request.

Decode share. The part of the time decoding lanes keep while a prompt prefills (--model-decode-share, 25 percent by default). See how ignis serves a request.

DFlash2. The 27B’s speculative-decoding drafter, grafted into its artifact: a five-layer, sliding-window model that proposes 1 to 7 tokens per round. See speculative decoding.

Draft. A token a drafter proposes ahead of the model, to be verified in the same pass. See speculative decoding.

Draft tokens. The most tokens a lane drafts per round (--spec-draft-tokens). See speculative decoding.

Drafter. Whatever proposes the drafts: DFlash2 for the 27B, the MTP head for Flash-Next. See speculative decoding.

Expert cache. On Flash-Next, the fixed-size pools of VRAM that hold the experts in use, replaced least-recently-used. It takes what the KV pool leaves and is refused below 12 GiB. See models and artifacts and the VRAM plan.

Family. The architecture an artifact runs as (Qwen3.8-27B, Qwen3.8-Flash-Next, Qwen-Image-2.1), read from the artifact’s header. Several artifacts can share one family. See models and artifacts.

Field. One setting, named the same as a flag, an IGNIS_* variable and a file key. See configuration.

Flash-Next. Qwen3.8-Flash-Next, the 125B mixture-of-experts model with 6B active parameters. See models and artifacts.

Generation cap. The most a request may generate, reasoning included: its max_tokens, or --model-default-max-tokens. See chat completions.

Group. The section a field belongs to (server, model, vram, reuse, switch, spec, vision, media, download, ngram, kv_disk). See configuration.

Headroom. The device memory a derived VRAM budget leaves to the desktop and other programs (--vram-headroom-bytes, 1,536 MiB by default). See the VRAM plan.

hq-e8-2b. The serving KV format: a compressed cache at 9,216 bytes per token on the 27B, against 65,536 for BF16. See the KV cache and prompt reuse.

Idle unload. Unloading a model after --model-idle-unload seconds with no request, while the server stays up; the next request loads it again. See model switching and idle unload.

Implicit switch. A switch a request triggers by naming another known model in model. It is controlled by --switch-allow-implicit. See model switching and idle unload.

Interactive. The default lane tag: a request from a person waiting on the answer. It ranks above agent. See Lane tag.

KV-disk. The tier below KV-RAM: a file per evicted state, so a sequence resumes where it stopped. On by default for Flash-Next. See KV-disk and the KV cache and prompt reuse.

KV page. A fixed block of 64 tokens of key and value cache, the unit the KV pool counts. See the KV cache and prompt reuse.

KV pool. The device memory for KV pages, shared by all lanes. See the VRAM plan and VRAM budget and the KV pool.

KV-RAM. The host-RAM tier of pinned memory that holds evicted sequences and retained state. It is held whole for the life of the load. See the KV cache and prompt reuse.

Lane. A concurrent decode slot. A request holds one while it generates; the 27B has eight, Flash-Next 1 to 8. See how ignis serves a request.

Lane tag. A request’s own statement of its class, interactive or agent, carried in the class field or an @<lane> suffix on the model name. It drives protection and eviction order. See how ignis serves a request.

Last switch. How the most recent switch or reload ended, kept on GET /v1/models. See model switching and idle unload.

Loop guard. The task that scores a request’s reasoning for repetition as it streams, and takes a persistent loop back. See reasoning and the loop guard.

MTP. Multi-token prediction: Flash-Next’s own draft head, shipped in a companion container. Off on the RTX 5090, on under the rtxpro6000 profile. See speculative decoding.

Nudge. The sentence in the model’s own voice that a rewind adds to the reasoning before it goes on. See reasoning and the loop guard.

Pinned hot set. Flash-Next’s calibration-hottest experts, kept in the expert cache for the life of the load with no copy in RAM (--vram-pinned-expert-bytes). See host RAM and profiles.

Prefill. Reading the prompt through the model once to fill its KV and state and draw the first token. See how ignis serves a request.

Prefill lane. The single slot where one prompt prefills at a time; others queue behind it. See how ignis serves a request.

Prefix reuse. Skipping the prefill of a prompt head an earlier request already read, by resuming from retained state that matches by content. See the KV cache and prompt reuse.

Profile. A named bundle of hardware-shaped defaults, below the config file and above the built-in defaults. See profiles.

Prompt checkpoint. The whole state of a request at its generation opener, kept after it ends so the next turn resumes there. See the KV cache and prompt reuse.

Provenance sidecar. See Sidecar.

Reasoning. The model’s thinking, written between <think> and </think> and returned as reasoning_content. See reasoning and the loop guard.

Restore floor. The prompt tokens a match in a lower tier must reuse beyond the best match above it to be worth the crossing. See the KV cache and prompt reuse.

Retained prefix. A shared prefix, ending at the system and tools block, kept alive with no request holding it so the next subagent can claim it. See the KV cache and prompt reuse.

Retained slot. A place reserved at load for one image of mutable state, on the device or in pinned host memory. See the KV cache and prompt reuse.

Retained state. Prompt checkpoints and retained prefixes: state no live request needs, first to go when room is needed. See the KV cache and prompt reuse.

Rewind. The loop guard’s action: cancel the request and continue it from the text before the loop, with a nudge. See reasoning and the loop guard.

Row budget. The most rows one Flash-Next MTP verify round may take across all lanes (--spec-draft-rows, 8 by default). See speculative decoding.

Scope. A model family’s own section of a field: qwen38 for the 27B, qwen38flashnext for Flash-Next. A value in a scope wins over the general one for that family. See configuration.

Served id. The name a client puts in model and GET /v1/models reports. It tells artifacts of one family apart. See models and artifacts.

Sidecar. The file beside an artifact (.graft.json or .conversion.json) that records its provenance, checked at load. See models and artifacts.

Speculation. Drafting tokens ahead and verifying them in one pass, so a round can yield several tokens. See speculative decoding.

Switch. Replacing the loaded model with another, as a full reload, without restarting the server. See model switching and idle unload.

Thinking budget. The most reasoning tokens before the server forces the close (--model-thinking-budget, 32,768 by default). See reasoning and the loop guard.

TTFT. Time to first token: how long a request waits before the first token. Prefix reuse and the decode share both move it. See the KV cache and prompt reuse.

Verify round. A decode round in which each lane carries its anchor and drafts and the model checks them all in one traversal. See speculative decoding.

VRAM budget. The device memory the whole process may hold, weights included: the free memory at start minus the headroom, or a named figure. See the VRAM plan.

VRAM plan. The layout of every byte of the budget, decided and printed at load; nothing is allocated on the device while serving. See the VRAM plan.

WDDM. The Windows display driver model, which pages device allocations to system RAM instead of failing them. See the VRAM plan.

YaRN. The rope-scaling method that stretches the checkpoint’s trained 262,144 positions to a longer context; chosen by auto from --model-max-context. See context, KV and speculation.