Grok Imagine Image 2.0 Is #2 in the World — And It Isn’t Free. Here’s the Free Alternative You Can Run Today

SpaceXAI — the lab formerly called xAI — shipped Grok Imagine Image 2.0 on August 7, 2026, and the numbers are real. It sits second in the world on both the Arena text-to-image board (1,320 Elo) and the image-edit board (1,439 Elo), behind OpenAI’s GPT-Image-2 on each. It does region-level editing with a magic wand and blends up to five reference images in one shot.

Here is the part the launch coverage buries: there is no free Grok Imagine Image 2.0. No open weights, no public API yet, and no free tier — xAI removed free image generation from X in January 2026 and from grok.com and the mobile apps in March. The entry price is SuperGrok at $30/month. So this post does what the launch posts won’t: it explains what actually shipped, then hands you a genuinely free open-weights model that runs on your own machine tonight.

At a Glance

• What it is: Grok Imagine Image 2.0, an editing-first image model from SpaceXAI, live as “Quality Mode” inside Grok on web, iOS and Android.
• What’s free: Nothing. Free image generation was pulled from every xAI surface between January and March 2026.
• What it costs: SuperGrok at $30/month is the entry point. The existing Grok Imagine Image API runs $0.02/image standard and $0.05/image quality at both 1K and 2K; an Image 2.0 API is listed as “coming soon.”
• The free path: Z-Image Turbo from Alibaba’s Tongyi lab — Apache-2.0, 8 sampling steps, and a live demo you can use right now with no account.
• The one catch: Z-Image Turbo is a generator, not an editor. If you specifically need magic-wand region editing, the free open stack doesn’t match Grok yet.

What Grok Imagine Image 2.0 Actually Ships

The headline feature is that editing is built in rather than bolted on. A Magic Wand tool changes only the region you point at and leaves everything else untouched. A segmentation tool picks precise areas. Background removal exports subjects with real transparency, which matters if you are handing files to a designer. Multi-Ref Editing accepts up to five input images in a single generation, so compositing that used to mean stitching sources by hand happens in one pass.

Smart Resize recomposes one image across nine aspect ratios, from a 1:2 tall banner to a 2:1 wide one, with the model filling in the new frame instead of stretching pixels. Preset templates cover product shots, headshots, e-commerce listings, icons, sprites, emotes and posters. SpaceXAI says training paid particular attention to typography and layout — the historic weak spot in dense, text-heavy compositions.

One honesty note on the benchmarks. The 1,439 and 1,320 Elo figures come from SpaceXAI’s own announcement citing public Arena leaderboards as of August 7, 2026 — they are not an independent benchmark study, and the gap to GPT-Image-2 (1,463 and 1,380) is narrow enough that it will move. Treat “second in the world” as a snapshot, not a fact about quality you will notice in your own work.

The Paywall Nobody Mentions

Grok Imagine used to have a free tier. It doesn’t. xAI restricted free image generation on X on January 9, 2026, then removed free access from grok.com and the Grok apps on March 19, 2026 — widely reported at the time and confirmed by multiple third-party trackers. Free Grok chat still exists; free Grok image generation does not, and there is no downloadable checkpoint at any price.

That’s a business decision, not a scandal. But for a creator on a budget it changes the question from “how good is Grok Imagine Image 2.0” to “what open model gets me close enough for free.” We’ve covered a few — Qwen-Image-Flash on a 12GB GPU, the 3GB Anima-2.9B anime build — but for general-purpose photorealism with text that renders, the answer is Z-Image Turbo.

Z-Image Turbo: The Free Model That Fills the Gap

Z-Image is a 6B-parameter single-stream diffusion transformer from Tongyi-MAI, Alibaba’s foundation model group. Turbo is the distilled variant: a finished image in 8 sampling steps instead of the base model’s 50, which is the difference between waiting and iterating. Alibaba reports sub-second generation on datacenter H800 cards; on consumer hardware expect a few seconds per 1024×1024 image.

It renders English and Chinese text inside images reliably enough for posters and mock layouts, follows multi-clause instructions well, and is licensed Apache-2.0 — commercial work, fine-tuning and redistribution all permitted. That matters, because plenty of “open” image models ship under research-only or region-restricted terms. The GitHub repo sits at 11.9k stars and recently added an MPS Flash Attention backend, so Apple Silicon is a supported target rather than a community hack.

Try It Free — Exact Links, Sizes and Commands

Zero-install, right now: open the official Z-Image Turbo demo on Hugging Face Spaces. It runs on free ZeroGPU, needs no account, and defaults to 8 steps at 1024×1024. Type a prompt, hit Generate, download. There’s a second community mirror at mrfakename/Z-Image-Turbo if the official one queues.

ComfyUI, the practical local route: grab the repackaged files from Comfy-Org/z_image_turbo. You need three pieces: a diffusion model, a text encoder, and the VAE. Pick by VRAM:

• 24GB card: z_image_turbo_bf16.safetensors (12.31 GB) + qwen_3_4b.safetensors (8.04 GB) + ae.safetensors (0.34 GB) → ComfyUI/models/diffusion_models, /text_encoders, /vae.
• 12–16GB card: z_image_turbo_int8_convrot.safetensors (6.20 GB) + qwen_3_4b_fp8_mixed.safetensors (5.63 GB) + the same VAE. About 12 GB total on disk.
• 8GB card: z_image_turbo_nvfp4.safetensors (4.51 GB) + qwen_3_4b_fp4_mixed.safetensors (3.48 GB) + VAE — roughly 8.3 GB total.
• GGUF, for llama.cpp-style backends: unsloth/Z-Image-Turbo-GGUF ships Q4_0 at 4.59 GB, Q3_K_M at 4.19 GB and Q2_K at 3.64 GB.

Python / Diffusers: the pipeline landed recently enough that you need Diffusers from source. Two commands:

pip install git+https://github.com/huggingface/diffusers

python -c "
import torch
from diffusers import ZImagePipeline
pipe = ZImagePipeline.from_pretrained(
    'Tongyi-MAI/Z-Image-Turbo',
    torch_dtype=torch.bfloat16,
    low_cpu_mem_usage=False,
).to('cuda')
img = pipe('a rain-soaked neon storefront at night, 35mm', num_inference_steps=8).images[0]
img.save('out.png')
"

One warning before you run that: the full Hugging Face repo is about 33 GB, because the transformer ships in fp32. If you only want to generate images, the ComfyUI split files above are the smaller, saner download.

The Honest Trade-Offs

Z-Image Turbo is not an editor. Tongyi ships a separate Z-Image-Edit checkpoint for instruction-based editing, but it runs 50 steps rather than 8 and offers no point-and-click region tools. If your job is retouching client photos, this is not a drop-in replacement for Grok.

Distillation costs you variety. Tongyi’s own comparison table rates Turbo’s output diversity as “Low” against the base model’s “Medium,” and Turbo drops classifier-free guidance entirely. Practical effect: reroll the same prompt ten times and you get ten similar images. For range, use the 50-step base Z-Image and accept the wait.

The leaderboard claim is vendor-run and stale. Z-Image’s model card cites Elo scores from Alibaba’s own AI Arena, last updated November 26, 2025 — a vendor benchmark on a vendor board, nine months old. It tells you the model is competitive, not that it beats anything specific today. Z-Image is a November 2025 release: the best free option, not the newest thing on the shelf.

Five Things to Do Differently

1. Leave steps at 8 and Time Shift at 3. Turbo is distilled for that exact schedule. Pushing to 20 or 30 steps burns time and usually degrades the image.
2. Quantize the text encoder before the transformer. The Qwen3-4B encoder is 8.04 GB — bigger than a quantized diffusion model. Dropping it to fp8 saves 2.4 GB with almost no visible cost.
3. Put poster text in the prompt, in quotes. Bilingual text rendering is Z-Image’s standout feature, and it works far better on a quoted exact string than on a description.
4. Generate at 1024, upscale after. Turbo was tuned around 1024-class resolutions. Asking for 2K directly costs quality; a separate upscale pass doesn’t.
5. Rent a paid editor by the month, not the year. Free generation plus one month of a paid editor when a client job needs region edits beats a $30 subscription running year-round.

Frequently Asked Questions

Is Grok Imagine Image 2.0 free?

No. xAI removed free image generation from X in January 2026 and from grok.com and the Grok apps in March 2026. Grok Imagine Image 2.0 requires a paid Grok plan, with SuperGrok at $30/month as the entry point. There is no free trial tier and no open-weights release.

Can I use Z-Image Turbo commercially?

Yes. Z-Image Turbo ships under Apache-2.0, which permits commercial use, modification, fine-tuning and redistribution with attribution. There are no region restrictions and no revenue thresholds. Confirm the licence on the model card before a large deployment, since labs occasionally revise terms.

What are Z-Image Turbo’s system requirements?

Alibaba states it fits comfortably within 16GB VRAM consumer devices at bf16. In practice: 8GB cards work with the nvfp4 build (roughly 8.3 GB of files), 12–16GB cards run the int8 build, and 24GB cards run full bf16. Apple Silicon is supported through the MPS Flash Attention backend in the GitHub repo.

Is there a free AI image editor with region-level control?

Not one that matches Grok’s magic wand. The closest free options are Z-Image-Edit for instruction-based edits, inpainting workflows in ComfyUI, and Google’s Gemini image tooling — we broke that one down in our Nano Banana Pro guide. For a local generator-plus-editor in one 4B package, Microsoft’s Mage-Flow is worth a look.

Does Grok Imagine Image 2.0 have an API?

Not yet. SpaceXAI lists a dedicated Image 2.0 API as “coming soon” with no date. The existing Grok Imagine Image API prices the standard tier at $0.02 per image and the quality tier at $0.05, flat per image at both 1K and 2K, regardless of prompt length.

We ran a head-to-head on this exact question over on the channel — Grok vs ChatGPT Text to AI Image Generator: One Is Better — and there’s more free-tool testing every week on the Creative AI Show channel.

Accurate as of August 20, 2026 — pricing, licences and leaderboard positions move fast. Benchmark figures attributed to SpaceXAI and Alibaba are vendor-reported unless stated otherwise.