FLUX.2 Klein 4B¶

Current status and review date¶

Decision snapshot

Product role: production-enabled named image-generation workflow; available=true and promoted=false describe the published workflow decision, not a model-family promotion or live worker state.
Selected or best-qualified configuration: FLUX.2 Klein 4B FP8 in ComfyUI v0.33.4, workflow image.flux2-klein-4b-fp8-v1, concurrency one, and fixed draft, standard, and high profiles.
Measured hardware: one NVIDIA GeForce RTX 5090 with 32,607 MiB in a Windows 11, Docker Desktop, and WSL2 media lane.
Evidence: direct functional/capacity qualification, exact-build and lifecycle acceptance, real Hermes/MCP calls, and six of eight strict bounded visual reviews passed.
Decision: enable only the exact named workflow at 512×512, 768×768, and 1024×1024, all with four steps and c1; reject caller-supplied graphs, models, dimensions, and substitutions.
Important limitation: concurrency above one, broad comparative image quality, seed repeatability, monotonic quality by tier, and a normal cold-start latency distribution were not tested.
Review dates: retained evidence through 2026-09-15; dossier reviewed 2026-09-17.

Review narrative¶

2026-09-15 — managed bringup smoke¶

One fixed high 1024×1024/four-step image job completed on the managed worker. The PNG passed stdlib integrity checks and one independent review, with a minor rough/pitted rather than brushed-steel mismatch. Its 38.335-second status delay is not image-generation latency. This did not alter workflow availability or promotion.

2026-08-28 — direct qualification¶

The exact managed workflow produced a decodable 258,472-byte PNG in 9.859 seconds and peaked at 12,919 MiB from a 943 MiB worker baseline. That initial cold c1 run established functional and capacity evidence only; it did not by itself establish routed behavior, broad visual quality, or a steady-state latency target.

2026-08-28 — gateway and exact-build validation¶

An isolated exact-build pass added two prompt-adherent PNGs and complete artifact controls. The live gateway pass then exercised cold approval, same-job resume, authenticated artifact delivery, real Hermes requests, and managed teardown while preserving the unrelated Qwen service.

2026-08-28 — production enablement and fix-forward closure¶

The production pass added one warm request per fixed profile and five cold draft regressions. Six of eight images passed strict independent review; two technically valid drafts failed origami/material fidelity and exact bird count. The controller, approval, cleanup, auth, and resume defects found along that path were fixed forward with retained regression evidence before the exact workflow became available.

Immutable identity¶

black-forest-labs/FLUX.2-klein-4b-fp8 revision 5b4408e59397a4a37ccb46afe426d8ed86379441, combined with the Qwen3 4B encoder and FLUX.2 VAE from Comfy-Org/flux2-klein revision 5f526678002e43af5551dadb73ce2e8c91b43afe. Exact file sizes and SHA-256 identities are recorded in the workflow bundle lock.

The immutable workflow descriptor records graph digest 991b63b8c61ff4322d72b8ae81ef43656f4905ddf4b0709c1989b84cfb8f2e4f.

Tested hardware and topology¶

One RTX 5090, 32,607 MiB, in a Windows 11 / Docker Desktop / WSL2 media lane. Direct qualification, isolated exact-build acceptance, and the production cross-host path with Primary Node gateway, Media Node resource ownership, and real Hermes on Companion Node were tested. The production router and bounded controllers changed; the unrelated Qwen service remained in place.

Engine, quantization, KV, context, and concurrency recipe¶

Exact workflow and runtime¶

ComfyUI v0.33.4 at 7a131a3a, CUDA 13.0, PyTorch 2.13.0+cu130, a digest-pinned base, curated node pins, and workflow image.flux2-klein-4b-fp8-v1 at graph digest 991b63b8...2e4f, c1. The server-owned profiles are draft 512×512, standard 768×768, and high 1024×1024, all four steps. KV cache and language-model context are not applicable to this diffusion workflow.

Reproduction boundary¶

The managed lifecycle is retained in examples/primary-node/serves.comfyui.toml and the exact models, node revisions, container identity, and workflow source are retained in the bundle lock. The initial direct qualification prompt text was intentionally not retained publicly; only its SHA-256 digest was published. Later production-quality prompts and dispositions are retained in the dated finding.

Evidence by measurement class¶

Functional and capacity¶

functional, capacity: the direct run produced a decodable 258,472-byte PNG in 9.859 seconds, peak 12,919 MiB from a 943 MiB worker baseline, with max queue running one and pending zero. The isolated exact-build pass added two prompt-adherent PNGs and complete artifact controls.

Routed client and bounded quality¶

The production pass added one real-Hermes warm request per fixed profile and five cold draft regressions. Six of eight samples passed strict independent review. Two technically valid draft results failed prompt adherence on origami/material fidelity and exact bird count. Warm gateway E2E measured 1.352/1.242/1.650 seconds for draft/standard/high.

Cold lifecycle¶

The 2,045.626-second cold E2E includes multiple fix-forward deployment failures and is not steady-state latency. The final server-issued exact-resume regression measured 908.936 seconds E2E and 0.087 seconds generation.

Decision and promotion state¶

Exact workflow availability¶

Enable the exact workflow at the three fixed c1 profiles. This is workflow availability, not a model-family promotion or a claim of broad image quality. Caller-supplied graphs, models, node paths, dimensions, and alternate generators remain unavailable.

Promotion boundary¶

available=true, promoted=false is the published decision. It does not authorize another workflow, route, model, host, or live worker mutation.

Failures and gotchas¶

Operational fix-forward history¶

The first qualification attempts exposed missing native build dependencies in the derived runtime. The production cold path later exposed controller timeout, Buildx, read-only client state, no-router scoping, approval recovery, descendant cleanup, and nested receipt defects; each was fixed forward with regression coverage. Later fixes moved live fleet probes into the router runtime, changed Hermes MCP auth to the least-authority router token, split the router-to-lifecycle and lifecycle-to-resource credentials, and made skill 1.0.4 copy the server-issued exact same-job resume bundle unchanged, keep nonterminal states in-turn, and preserve the current request language.

Post-run revision 5ea1edc and skill 1.0.5 then closed the profileless empty-profile boundary, the five-input/seven-field replay distinction, and the malformed-bundle cancellation exception without changing the measured image runtime.

Quality and generalization limits¶

One cold c1 run is not a throughput distribution or latency target. Six strict passes and two retained draft failures do not establish broad text rendering, hands, counting, material fidelity, seed repeatability, or monotonic quality by tier. Those properties and concurrency above one were not tested.