Krea 3 and EmbeddingGemma 2 speech-to-text: two smaller AI releases worth knowing

Two releases from the week of October 12, 2026, drew more developer attention than their size suggests. Krea 3, launched October 15, is one of the image models built for editing at scale rather than generating from scratch. And a developer turned EmbeddingGemma 2, Google's small open-weight embedding model, into an on-device speech-to-text app. One is a commercial API; the other is a small hack on an open model.

This story is based on Krea's announcement, product page and API documentation, fal's Krea 3 page, Google's developer blog and model card, the Embedding Gemma Transcribe project page and its Hugging Face port, checked October 16, 2026. On Hacker News, Krea 3 drew nearly 300 points and the EmbeddingGemma 2 project more than 400.

The short version

• Krea 3: an image-editing model from Krea. $0.03 per 1K image, $0.06 per 2K image via Krea's API, which is in early access. Krea says most edits take under 8 seconds. No open weights; Krea offers self-hosting on enterprise plans.
• EmbeddingGemma 2: Google's 740M-parameter, Apache 2.0 model that puts text, images, video and audio into one vector space.
• The speech-to-text hack: Shahir's free Mac app keeps EmbeddingGemma 2's audio encoder frozen and trains a small "reader" that picks the nearest of about 229,000 English words. About 5% word error rate, per his page.

What is Krea 3?

Krea, the AI image and video platform, announced Krea 3 on October 15, 2026. It is aimed at edits: change one element, such as a sofa, and leave the rest of the scene alone, including product labels and logos. Krea says it keeps people consistent across images, works on photos, paintings, vector art and ink drawings, and generates natively at 2K resolution.

Krea's headline claim is speed. It says most edits finish in under 8 seconds at 2K, with the fastest setup at about 4 to 5 seconds. CEO Víctor Perez says Krea 3 "removes the tradeoff between speed and quality."

That speed comparison is vendor-reported. Krea measured median time per edit on its own platform from September 24 to October 1, 2026, queue time included, and compared it with 33 seconds for Nano Banana Pro, 27 seconds for Nano Banana 2 and 90 seconds for GPT Image 2. No independent evaluator has published results.

How much does Krea 3 cost, and where can you use it?

Last verified: October 16, 2026 (Krea 3 API overview)

Reference fees add up: an edit with several 2K references costs several times the headline $0.06.

• Krea app: available on krea.ai.
• API: early access only. Krea's docs say only workspaces in the Krea 3 beta can call it; others contact sales.
• Partners: Krea's blog says Krea 3 is live through fal, Hedra, Together AI, Nous Research, OpenRouter and Replicate. fal's page lists it as live on its API and playground, billed at $0.025 per megapixel for the output image plus each reference image, a different pricing scheme from Krea's own.
• Reference images: sources disagree. Krea's blog says up to 6 per edit, while its product page, its API docs and fal's page say up to 8 (Krea's API allows 4 in zero-data-retention workspaces). Check the platform you use.

Is Krea 3 open source?

No. Krea's announcement and API docs name no license and mention no weights release. Its product page says self-hosting is available "on enterprise plans," running on the customer's own GPUs under license terms negotiated with Krea. Treat it as a commercial model. If you need an editing model you can run yourself, our coverage of FLUX 3 Image Klein explains what Black Forest Labs' open release allows and doesn't.

What is EmbeddingGemma 2?

EmbeddingGemma 2 is an embedding model from Google DeepMind, announced on Google's developer blog on October 6, 2026. It turns text (including code), images, video frames and audio into vectors in one shared space, so a spoken question can retrieve a matching photo or document.

According to Google's model card:

• Size: 740M parameters: a 270M text model, a 170M vision encoder and a 300M audio encoder.
• Output: 768-dimensional vectors, which can be truncated to 512, 256 or 128 dimensions.
• Context: 8,192 tokens.
• License: Apache 2.0 on Hugging Face.

Google says the full multimodal model uses about 567MB of active RAM on a Pixel 11 Pro. Google's own benchmark table, which is vendor-reported, shows a small gain over the first EmbeddingGemma on multilingual text retrieval (MTEB multilingual v2: 61.36 vs. 61.15) and a larger one on code (78.68 vs. 68.76).

Who turned EmbeddingGemma 2 into speech-to-text, and how?

An embedding model isn't meant to transcribe anything. It outputs vectors, not words. The "We turned EmbeddingGemma 2 into a speech-to-text model" post on Hacker News (October 13) linked to Embedding Gemma Transcribe, an independent project that says it is "not affiliated with or endorsed by Google." In the thread, a developer who introduced himself as Shahir said he built it.

The trick, as the project page describes it:

1. EmbeddingGemma 2's audio encoder stays frozen.
2. A small trained "reader" maps short slices of audio into the model's shared space.
3. Because text lives in that same space, each slice is matched to the closest of about 229,000 English words, using EmbeddingGemma's own text embeddings for those words.
4. A tiny extra model adds punctuation and capitals.

Only the reader was trained, in under two hours on eight H100 GPUs using openly licensed English speech, according to the page. The app is free, runs on Apple Silicon Macs (macOS 14 or later) on the Mac's Neural Engine, and downloads 715MB of models on first launch. The page says "your audio never leaves the machine."

The project reports about 5% word error rate, versus 2–3% for leading speech models, and 288 times real time on a base M3 MacBook Air. These are the developer's own numbers. The page says only its own two rows were measured on that MacBook Air; the rows for other models come from FluidAudio's published benchmarks on newer Macs with different audio, "a rough guide, not a like-for-like ranking."

A separate open port by Hugging Face user mlboydaisuke, an independent reproduction of Shahir's method, publishes more detail. Its current reader (v2) has 19.6M parameters, was trained on LibriSpeech's 960 hours, and reports a 4.21% word error rate on LibriSpeech test-clean and 10.42% on test-other, without a language model. Its first reader, trained on the smaller train-clean-100 set, reported 6.81% and 15.10%. That port's bundles are Apache 2.0, and deployments must follow Google's Gemma Prohibited Use Policy.

Both are English only, with no measured results on noisy or conversational speech. Shahir wrote on Hacker News that he plans to open-source the project "soon," and is working on a multilingual reader. As of October 16, the app's source isn't public.

Why these matter

Krea 3 is a price signal. If Krea's numbers hold up, a 2K edit at $0.06 is less than half Krea's quoted list price for Nano Banana Pro ($0.135), which pushes editing toward high-volume uses like product catalogs.

The EmbeddingGemma 2 project is a capability signal. A shared multimodal space turned out to be good enough to transcribe speech with a tiny trained add-on. That strengthens the case for small local models, which our guides to small models for laptops and Liquid AI's LFM2.5 cover.

What's confirmed, reported and uncertain

• Confirmed (official pages): Krea 3's launch date, API prices, early-access API status and enterprise-only self-hosting; EmbeddingGemma 2's size, license, dimensions and context length.
• Vendor- or developer-reported: Krea's speed and price comparisons; Google's benchmark scores and RAM figures; Embedding Gemma Transcribe's error rate and speed; the port's LibriSpeech results.
• Uncertain: whether Krea will ever release Krea 3 weights publicly; the exact reference-image limit on each platform; when Shahir's code will be published, and under what license.