Repository navigation
Image model: Qwen-Image-2.1 + fast variants (tier 1) — the currently-trending image family (3M+/mo wave), upstream flagship parity #101
Description
Parent: #95 (tier 1, follows the small+fast tier) · Prerequisite: #91 (mflux seam). The 2026-10-06 hot-list refresh found this is the currently-trending image model family.
Model card
- Repo:
Qwen/Qwen-Image-2.1— released 2026-09-14 (3 weeks old), 94.5k dl/mo official + 3k likes already; the derivative wave is enormous: an uncensored GGUF at 1.64M dl/mo, unsloth GGUF 612k, Viggle turbo distill 287k, Pruna 206k. Licenseother(verify the Qwen license terms before defaulting to it). - Upstream parity: ✅✅ — upstream vllm-omni carries Qwen-Image recipes (Qwen-Image, -Edit, -2512) as their flagship t2i examples; 2.1 is the current generation.
- Params: 7.1B single-stream block-causal DiT + separate 8B Qwen3-VL text encoder; causal VAE with 16× spatial compression. 40-step guidance-free sampling by default.
- MLX: first-class mflux support (
mflux-generate-qwen-2.1,-edit) — t2i, strength-img2img, multi-reference editing with prefix KV caching, LoRA, RGBA output. - Memory: DiT @4-bit ≈ 3.6 GB + TE @4-bit ≈ 4.3 GB ≈ ~8 GB total → solo-qualified only on the 10.7 GB working set; no mixed residency.
- Disk: ~15 GB unquantized → quantize at download; another reason it follows the small tier.
Work
- License verification first (record verdict here before any download)
- mflux seam backend branch; defaults: 40-step guidance-free (measure before fixing defaults)
- Fast-variant decision lives here: Viggle turbo (7.1B total, few-step, 287k/mo, license
other) vs Pruna (206k) vs staying official-base — record verdict, then serve the winner - Reference-editing path (prefix KV) as the follow-on after t2i — second edit-native model after Image model: FLUX.2-klein-4B (tier 1) — 394k/mo, 4-step, Apache-2.0, smallest of the category, edit-native #100
- Weight-gated e2e + output gates; latency/memory table (solo; time-to-image @1024² 40-step and fast-variant steps)
- README models-table row
Expectations
7.1B × 40 steps ≈ 6× the compute of Z-Image-Turbo's 9 → expect ~10–30 s per image official-base; single-digit seconds if the turbo variant wins the decision. The value here is prompt understanding/world knowledge + upstream parity, not latency.
Acceptance
License + fast-variant verdicts recorded first; weight-gated e2e green; latency table in the PR; README row added.
Activity
hsliuustc0106 commentedon Oct 7, 2026
Verdicts (recorded first, per the acceptance order)
License — non-commercial.
Qwen/Qwen-Image-2.1ships the Qwen Research License Agreement (2026-09-20): grant of rights is explicitly
"FOR NON-COMMERCIAL PURPOSES ONLY" (research/evaluation); commercial use requires a separate
license from Qwen. Redistribution obligations: pass the agreement through, keep the
attribution notice. Verified from the LICENSE file in the repo (not from the tag).
Both fast variants inherit the same agreement —
Viggle/Qwen-Image-2.1-viggle-turbo'sLICENSE and
PrunaAI/Pruna-Qwen-Image-2.1's LICENSE are the identical Qwen Research text(distillations are derivatives under §3). → Served with the license printed at load and
flagged in README/docs; the Apache-2.0 tier (#99/#100) is the commercial-safe path on the
same seam.
Fast variant — Viggle turbo wins. mflux (0.21.0) carries first-class support: the
distilled LoRA + a built-in
ViggleTurboSchedulerthat samples the LoRA's 6 trained sigmanodes (6 steps, no CFG — its docstring cites the v0.2.1 model card). Pruna ships merged
dense transformers (
p_qwen_image_2.1_5step_v0.1.safetensors) with no mflux loader path→ not selected; recorded as possible follow-up (custom-transformer load). Staying
official-base is the quality reference; the turbo LoRA is the fast lane. A/B'd v0.3 vs
v0.2.1 r128 at 512²/seed 1000: v0.3 ghosts less → v0.3 default.
Weights path.
mlx-community/Qwen-Image-2.1-mflux-q4— pre-quantized mflux-format(4-bit DiT + 4-bit TE + VAE, 9.6 GB on disk), loads in ~8 s with 8.93 GiB resident;
no dense download/quantize pass needed (dense path still available via
--quantize).Delivered in #109
/v1/images/generations) — thatseam had not landed yet, so it lands here with Qwen-Image-2.1 as its first backend;
Image generation engine: /v1/images/generations on MLX — mflux seam, [image] extra, time-to-image doctrine #91's Z-Image-Turbo reference-path acceptance stays open for Image model: Z-Image-Turbo 6B (tier 1, engine reference path candidate) — 592k/mo, 9-step, Apache-2.0, mflux-native #99.
--image-lora org/repo:file --image-scheduler viggle_turbo(steps=6 enforced); one-shot
vllm-omni-mlx image.base 40-step across 512²→1328², turbo 6-step at 512²/1024², per-step cadence, MLX peak +
RSS — in the PR and docs/image.md.
issue's acceptance cell (valid 1024² image, default 40-step path). README row added.
Remaining from the checklist: reference-editing path (prefix KV) — follow-on PR after
this t2i lands (second edit-native model after #100).