DEV Community: HexGrid Cloud

Can Krea 2 Turbo Create Product Ads? I Tried 5 Text-to-Image Experiments

simran — Thu, 29 Oct 2026 15:46:35 +0000

Introduction

Krea 2 Turbo offers an interesting way to explore these aspects through text-to-image generation. Using ComfyUI, five product-advertising examples demonstrate how the model interprets different prompts and translates creative directions into advertising visuals.

This article covers the prompts used for each product, the generated results, and the creative decisions behind them. Readers will get a better understanding of how to structure prompts for different advertising styles, guide product placement and lighting, and use Krea 2 Turbo to create visuals suited to different commercial applications.

Model Configuration and Inference Optimizations

For our image generation experiments, we used Krea 2 Turbo in ComfyUI, combining an INT8-quantized diffusion model with an FP8 text encoder to reduce memory requirements and make inference more efficient.

Our generation setup consisted of:

Component	Configuration
Diffusion Model	Krea 2 Turbo — INT8 ConvRot
Text Encoder	Qwen3-VL 4B — FP8 Scaled
VAE	Qwen Image VAE
Inference Framework	ComfyUI

Why These Optimizations Matter

The INT8 ConvRot diffusion model reduces model-weight memory requirements compared with higher-precision variants, making Krea 2 Turbo more practical to run on GPUs with limited VRAM.

The Qwen3-VL 4B FP8 Scaled text encoder further reduces the memory footprint of prompt processing, while the Qwen Image VAE handles image encoding and decoding.

Together, these components provide a memory-efficient configuration for exploring Krea 2 Turbo's image generation capabilities in ComfyUI. Actual generation speed, VRAM usage, and output quality depend on the GPU, resolution, and sampling settings.

Prompting Approach

For product ads, describe the object's physical construction and finish, then decide where it sits in the frame. A single hero product is usually easier to art-direct than a crowded scene. Keep the headline short, put exact copy in quotation marks, and specify a clear text area. If exact packaging, logos, prices, or legal copy matter, add them in a design editor after generation.

1. Premium Headphones - Minimalist Luxury

Format: 4:5 portrait
Creative direction: Warm ivory, matte black, brushed metal, soft studio light.

Prompt

Create a premium commercial advertisement for fictional VANTA wireless over-ear headphones. Show one pair of matte-black headphones with brushed-aluminum hinges and soft memory-foam ear cushions, floating at a subtle three-quarter angle above a warm ivory podium. Place the product in the right two-thirds with a controlled soft shadow. Use large-format studio photography, an 85mm lens, soft directional light from the upper left, realistic metal reflections, and a warm-gray and charcoal palette. Leave the left third clean for typography. Add the exact bold black headline "TUNE OUT THE NOISE", the smaller line "Sound, on your terms." and a minimal CTA reading "EXPLORE VANTA". Keep the spacing sophisticated and the product geometry precise. No other writing or decorative objects.

What to evaluate: Headphone geometry, reflections, silhouette clarity, and whether the left-side copy area remains uncluttered.

Output Image

2. Reusable Water Bottle - Performance Advertising

Format: 1:1 square
Creative direction: Electric cobalt, orange contrast, frozen water motion.

Prompt

Create a high-energy advertisement for fictional reusable-bottle brand ROVE. A large brushed stainless-steel bottle with an electric-cobalt finish and a simple screw cap is suspended diagonally in mid-air, surrounded by crisp water splashes and translucent droplets. Use a saturated orange studio background, high-speed flash photography, believable reflections, and dynamic but controlled motion. Keep the bottle as the unmistakable hero. At the top, add the exact white headline "TAKE IT EVERYWHERE", followed by "Meet your everyday bottle." At the bottom, place a clean white CTA reading "SHOP ROVE". Use modern athletic-brand art direction, strong color contrast, and an uncluttered composition. No extra products or text.

What to evaluate: Whether the bottle remains legible against the energetic background and whether water effects support rather than obscure the product.

Output Image

3. Designer Desk Lamp - Premium Homeware

Format: 16:9 landscape
Creative direction: Warm lamp glow against cool twilight.

Prompt

Create an elegant wide-format advertisement for fictional lighting brand FORM. A sculptural mushroom-shaped brushed-aluminum table lamp sits on a dark walnut desk in a contemporary minimalist apartment at twilight. The lamp casts a warm pool of light over a small open notebook while cool blue evening light enters through a large window. Use architectural lifestyle photography, realistic material detail, and restrained luxury styling. Keep the lamp and interior on the right half; leave the left half as a clean, shadowed wall. Add the exact off-white serif headline "LIGHT CHANGES EVERYTHING", the small uppercase line "OBJECTS FOR LIVING", and the CTA "DISCOVER FORM". No other text.

What to evaluate: The balance between warm and cool light, believable illumination, and negative space for the headline.

Output Image

4. Travel Luggage - Lifestyle Conversion Ad

Format: 4:5 portrait
Creative direction: Sage-green luggage, sunlit modern station, aspirational travel.

Prompt

Create a premium travel-lifestyle advertisement for fictional AERA carry-on luggage. Show a muted sage-green hard-shell suitcase with vertical grooves and polished black wheels standing beside a traveler wearing cream tailored trousers and white sneakers; show only the traveler's lower body. Set the scene in a sunlit modern European train station with pale stone architecture and soft distant motion blur. Use natural morning light, editorial 35mm photography, subtle film grain, and quiet-luxury color grading. Keep the suitcase fully visible in the lower center and leave a softly blurred cream area above it. Add the headline "GO SOMEWHERE NEW", the smaller line "Travel looks good on you." and the CTA "MEET AERA". No unrelated typography.

What to evaluate: Suitcase proportions, wheel and handle consistency, and whether the background supports the product without distracting from it.

Output Image

5. Portable Power Bank - Bold Technology Launch

Format: 4:5 portrait
Creative direction: Obsidian black, electric cyan, precise 3D materials.

Prompt

Create a futuristic product advertisement for fictional PULSE portable power bank. Show one precisely shaped rectangular black power bank with rounded edges, a subtle cyan indicator light, and a braided charging cable floating above a glossy obsidian platform. Use translucent glass panels with controlled cyan illumination for depth, ray-traced reflections, crisp rim lighting, and a rich black-and-cyan palette. Center the product slightly below the middle. Add the bold geometric headline "POWER IN YOUR POCKET", the supporting line "Built for the everyday." and a cyan CTA reading "DISCOVER PULSE". Keep exactly one main device and avoid random numbers or extra text.

What to evaluate: Device geometry, cable plausibility, reflections, and the visual hierarchy between product and headline.

Output Image

Krea 2 Turbo on HexGrid.cloud

Run Krea 2 Turbo *on HexGrid in one click — no CUDA setup, no dependency hunting, no environment debugging.*

The template comes preconfigured with ComfyUI, the required runtime, and the Krea 2 Turbo setup so you can launch a GPU workspace and start generating immediately.

No manual CUDA installation. No PyTorch version matching. No broken requirements. No hours spent wiring the environment together.

Just choose the Krea 2 Turbo template, start the instance, and open ComfyUI.

Pricing starts from $0.58/hour*, so you can spin up a GPU only when you need it and shut it down when you’re done.*

Testing FastH3 Omni in ComfyUI: 5 Video Experiments on an NVIDIA L40

simran — Tue, 27 Oct 2026 18:35:23 +0000

Introduction

Getting an AI video model to generate a good-looking clip is one thing. Getting it to keep the same character, preserve product details, and follow a scene across multiple shots is a different challenge.

I wanted to see how FastVideo FastH3 Omni Reference handles these situations in practice, so I ran five experiments in ComfyUI on an NVIDIA L40 with 48 GB of VRAM. The tests cover everything from a short fashion film and a luxury product advertisement to character consistency, a sci-fi scene, and a multi-shot commercial.

Rather than judging the model by a single output, I used different reference images and prompts to see how it handles identity, clothing, objects, camera movement, and scene composition. I've included the prompts, generation settings, and resulting videos so you can try similar experiments yourself.

Model Configuration and Inference Optimizations

For our five product advertising experiments, we used the FastVideo FastH3 OmniRef 8-step model in ComfyUI, with quantized model weights and optimized attention to make reference-to-video generation more memory-efficient.

Our generation setup consisted of:

Component	Configuration
Diffusion Model	FastH3 OmniRef 8-Step Pruned INT8 ConvRot
Text Encoder	Qwen3-VL 32B MiniMax H3 (NVFP4 AWQ)
Video VAE	MiniMax H3 INT8 ConvRot
Audio VAE	MiniMax H3 FP32
Video Sampling Shift	12.0
Audio Sampling Shift	3.0
Attention Backend	ComfyKitchen Attention
Sparse Attention	VSA (Video Sparse Attention)
VSA Keep Percentage	10%
VSA Active Range	0.0–1.0 (entire sampling schedule)

Why These Optimizations Matter

The INT8 ConvRot diffusion model and video VAE reduce model-weight memory requirements compared with their higher-precision counterparts, while the NVFP4 AWQ text encoder further reduces the memory footprint of the large Qwen3-VL 32B encoder.

We also configured ComfyKitchen Attention alongside VSA with a 10% keep ratio, enabling sparse attention throughout the sampling schedule. This is intended to reduce attention computation and improve inference efficiency, although the actual performance gains depend on the hardware and workload.

Combined with FastH3's 8-step distilled architecture, these optimizations provide a practical configuration for experimenting with high-quality, reference-guided product advertisements without relying on the full-precision MiniMax H3 pipeline.

Generation 1: Streetwear Fashion Film — 5 seconds

Input Image 1

Input Image 2

Input Image 3

Prompt

subject_definitions: <Subject 1> is the fictional adult fashion model in <Picture 1>. <Picture 2> is the specific oversized black puffer jacket that <Subject 1> wears. <Picture 3> is the brutalist concrete hall used as the location.

summary: [reference generation] A five-second, single-take high-fashion campaign with <Subject 1> wearing <Picture 2> inside <Picture 3>; powerful, minimal, photorealistic.

retention_analysis: <Subject 1> (Shot 1): fully_preserved — preserve the face, black choppy hair, facial proportions, relaxed androgynous styling, dark cargo trousers and boots of <Picture 1>. Preserve the puffer jacket's bulky silhouette, horizontal red chest band, quilting, collar and zipper exactly as <Picture 2>. Preserve the bare concrete and tall windows of <Picture 3>. Do not copy the gray studio background from <Picture 1> or <Picture 2>.

detailed_description: [Shot 1, 00:00–00:05] One continuous live-action full-body fashion shot. <Subject 1> stands at the far end of the hall from <Picture 3>, wearing the jacket from <Picture 2> over the outfit from <Picture 1>. The model takes three slow deliberate steps toward the camera. The camera tracks backward smoothly at hip height, then slowly raises to chest height. The red band stays a solid physical garment detail, not a floating graphic. At the end the model stops and turns their head ten degrees toward the windows. Natural cloth compression and soft shoe contact, realistic restrained movement, cold overcast light, subtle reflections on the concrete floor. No cuts, no extra people, no logos, no text overlays.

overall_soundscape: A low airy interior room tone, three distinct heavy rubber-sole footsteps, quiet fabric rustling and a distant wind through the windows.

non_diegetic_music: Sparse industrial sub-bass pulses with spacious pauses; understated, not overpowering the footsteps.

Setup Configs

Red Band — fashion film Duration: 5 seconds; choose 124 frames / 24 fps. Suggested FIRST test output: about 1-megapixels, preserve reference aspect ratio via the official template if possible. Ref_image_size: match. Steps: 8 | guidance: 1.0 | sigma shifts: video 12.0 / audio 3.0 | VSA: 10%

Connect ordered inputs into ref_images on MiniMaxH3ReferenceToVideo: <Picture 1> -> Picture_1.png -> the protagonist’s face, haircut, base pants and boots <Picture 2> -> Picture_2.png -> the exact black quilted puffer jacket with its broad red horizontal panel <Picture 3> -> Picture_3.png -> the concrete windowed hall and its cold daylight

Do not connect these images to the First Last Frame to Video node.

Actual render quality and VRAM requirements will vary.

Output

Generation 2: Luxury Product Advertisement — 5 seconds

Input Image 1

Input Image 2

Input Image 3

Prompt

subject_definitions: <Subject 1> is the fictional adult woman from <Picture 1>. <Object 1> is the exact sculptural chrome wraparound sunglasses from <Picture 2>. <Location 1> is the curved concrete atrium from <Picture 3>.

summary: [reference generation] A five-second premium eyewear commercial: <Subject 1> models <Object 1> in <Location 1> with elegant natural movement and product clarity.

retention_analysis: <Subject 1> (Shot 1): fully_preserved — preserve the facial proportions, center-parted black hair, skin tone and black sleeveless top of <Picture 1>. <Object 1> (Shot 1): fully_preserved — preserve the unusual sweeping chrome top line, dark lens shape and bright polished metallic surface of <Picture 2>; it must rest realistically on her nose and ears. <Location 1> (Shot 1): background_reference — use the circular skylight, pale concrete curves, sunlight geometry and empty architectural volume from <Picture 3>. Do not treat the gray product-photo background as the scene.

detailed_description: [Shot 1, 00:00–00:05] One continuous photorealistic 85mm portrait shot inside <Location 1>. Medium close-up of <Subject 1> already wearing <Object 1>, standing almost still near a pool of skylight. A very slow camera arc, no more than 20 degrees, moves from her three-quarter left to a more frontal angle. She gently turns her chin toward the light and adjusts the right temple of her sunglasses once with her right hand. The chrome frame catches one clean highlight moving across it; keep the sunglasses design geometrically stable. Realistic hand-to-frame contact, natural eyelid motion, fine skin detail, high-end editorial color. No abrupt cuts or added products.

overall_soundscape: Open architectural ambience, quiet distant footsteps with gentle reverberation, soft clothing movement and a very faint breeze.

non_diegetic_music: Three restrained crystalline electronic notes over a warm, minimal ambient pad.

Setup Configs

Chrome eyewear — model, product and architecture Duration: 5 seconds; choose 124 frames / 24 fps. Suggested FIRST test output: about 1 megapixels, preserve reference aspect ratio via the official template if possible. Ref_image_size: match. Steps: 8 | guidance: 1.0 | sigma shifts: video 12.0 / audio 3.0 | VSA: 10%

Connect ordered inputs into ref_images on MiniMaxH3ReferenceToVideo: <Picture 1> -> Picture_1.png -> the model’s facial identity and dark hair <Picture 2> -> Picture_2.png -> the silver sculptural wraparound frame and dark lenses <Picture 3> -> Picture_3.png -> the circular skylight and sculptural concrete atrium

Do not connect these images to the First Last Frame to Video node.

Actual render quality and VRAM requirements will vary.

Output

Generation 3: Character Consistency — 5 seconds

Input Image 1

Input Image 2

Input Image 3

Prompt

subject_definitions: <Subject 1> is the fictional woman in <Picture 1>. <Garment 1> is the black oversized puffer jacket with one broad solid red band in <Picture 2>. <Location 1> is the bright circular concrete atrium in <Picture 3>.

summary: [reference generation] A five-second wardrobe-transfer test: keep the identity of <Subject 1>, dress her in <Garment 1> and place her in <Location 1> in a new shot.

retention_analysis: <Subject 1> (Shot 1): fully_preserved — use the facial features, long center-parted black hair and skin appearance from <Picture 1>, even though the original outfit and gray studio are not used. <Garment 1> (Shot 1): fully_preserved — transfer the shape, thick quilting, upturned collar, black glossy fabric and red chest stripe from <Picture 2>. <Location 1> (Shot 1): background_reference — match the round skylight, raw concrete and sunlit floor from <Picture 3>. Do not transfer the head or body of a mannequin or an invisible product model.

detailed_description: [Shot 1, 00:00–00:05] Single unbroken live-action fashion shot. <Subject 1> stands under the skylight in <Location 1> wearing <Garment 1>, zipped up to the collar. Camera begins at waist-height in a medium shot and performs a gentle slow push-in. She looks up briefly toward the skylight, then lowers her eyes to the camera with a subtle serious expression. She slips both hands into the jacket pockets. The red band remains aligned and fixed around the jacket as her posture changes; jacket folds move naturally. Strong architectural shadows, realistic skin and garment materials. No cut, no extra clothing, no text or logos.

overall_soundscape: Bright quiet hall ambience, a light sleeve rustle, soft echo of distant air circulation.

non_diegetic_music: Slow minimal analog synthesizer and a low cinematic pulse.

Setup Configs

Cross-reference wardrobe transfer Duration: 5 seconds; choose 124 frames / 24 fps. Suggested FIRST test output: about 1 megapixels, preserve reference aspect ratio via the official template if possible. Ref_image_size: match. Steps: 8 | guidance: 1.0 | sigma shifts: video 12.0 / audio 3.0 | VSA: 10%

Connect ordered inputs into ref_images on MiniMaxH3ReferenceToVideo: <Picture 1> -> Picture_1.png -> face and person identity from collection B <Picture 2> -> Picture_2.png -> jacket and design from collection A <Picture 3> -> Picture_3.png -> new location, different from both source photos

Do not connect these images to the First Last Frame to Video node.

Actual render quality and VRAM requirements will vary.

Output

https://www.youtube.com/shorts/tgua6_rCpkk

Generation 4: Sci-Fi Cinematic Scene — 10 seconds

Input Image 1

Input Image 2

Input Image 3

Prompt

subject_definitions: <Subject 1> is the fictional adult with short dark hair, a fitted dark long-sleeve top and black cargo trousers in <Picture 1>. <Subject 2> is the fictional adult woman with long center-parted black hair and black sleeveless top in <Picture 2>. <Location 1> is the empty concrete hall with tall windows from <Picture 3>.

summary: [reference generation] A restrained ten-second science-fiction mystery conversation between <Subject 1> and <Subject 2> in <Location 1>. Each character keeps their own face, hairstyle and outfit.

retention_analysis: <Subject 1> (Shot 1): fully_preserved — short tousled black hair, angular face, dark fitted top and cargo trousers from <Picture 1>. <Subject 2> (Shot 1): fully_preserved — long center-parted hair, distinct softer face and black sleeveless top from <Picture 2>. Keep the two characters visually distinct; do not blend facial features or swap their voices. <Location 1>: background_reference — retain the empty concrete geometry and diffuse daylight from <Picture 3>.

detailed_description: [Shot 1, 00:00–00:10] One continuous, steady, cinematic medium two-shot in <Location 1>. <Subject 1> stands on screen left, facing <Subject 2> on screen right, separated by one meter. Slow lateral camera drift. At 00:01 <Subject 1> looks to <Subject 2> and says quietly, [English] Did you hear that? A small pause. <Subject 2> looks toward the upper windows and replies at 00:05, [English] It came from upstairs. They exchange a brief, concerned glance as the camera settles. Subtle expression changes and synchronized lips, natural conversational pacing. Keep both faces visible, correct speaker-to-mouth alignment, no cuts and no other characters.

overall_soundscape: Two distinct natural speaking voices, long concrete-room reverberation, very distant low mechanical hum, faint wind through the window joints.

non_diegetic_music: N/A

Setup Configs

Two-character dialogue scene Duration: 10 seconds; choose 243 frames / 24 fps. Suggested FIRST test output: about 1 megapixels, preserve reference aspect ratio via the official template if possible. Ref_image_size: match. Steps: 8 | guidance: 1.0 | sigma shifts: video 12.0 / audio 3.0 | VSA: 10%

Connect ordered inputs into ref_images on MiniMaxH3ReferenceToVideo: <Picture 1> -> Picture_1.png -> Character A, short dark haircut and black top <Picture 2> -> Picture_2.png -> Character B, long center-parted dark hair <Picture 3> -> Picture_3.png -> quiet concrete hall with windows

Do not connect these images to the First Last Frame to Video node.

Actual render quality and VRAM requirements will vary.

Output

Generation 5: Multi-Shot Commercial — 10 seconds

Input Image 1

Input Image 2

Input Image 3

Input Image 4

Input Image 5

Prompt

subject_definitions: <Subject 1> is the fictional adult with short dark tousled hair from <Picture 1>. <Subject 2> is the fictional woman with long center-parted black hair from <Picture 2>. <Product 1> is the oversized black quilted puffer with a bold solid red chest band from <Picture 3>. <Product 2> is the sculptural chrome sunglasses with dark lenses from <Picture 4>. <Location 1> is the circular concrete atrium in <Picture 5>.

summary: [reference generation] A ten-second three-shot luxury capsule-collection film connecting two models and two hero products in one architectural location. Stylish, restrained, consistent and photorealistic.

retention_analysis: <Subject 1> appears in Shot 1 and Shot 3: fully_preserved — same short haircut and face as <Picture 1>. <Subject 2> appears in Shot 2 and Shot 3: fully_preserved — same long black hair and face as <Picture 2>. <Product 1> appears in Shots 1 and 3: fully_preserved — exact black puffed silhouette and red horizontal stripe from <Picture 3>. <Product 2> appears in Shots 2 and 3: fully_preserved — stable polished chrome geometry and dark lenses from <Picture 4>. <Location 1> appears in all shots: background_reference — curved concrete, geometric daylight and central skylight from <Picture 5>.

detailed_description: Cinematic live-action advertising film with clean physical cuts; no superimposed typography or brand marks. [Shot 1, 00:00–00:03.3] Low-angle medium full-body shot of <Subject 1> in the atrium wearing <Product 1>. One confident step, a subtle camera push, and a clean view of the broad red stripe. Sunlight grazes the quilted surface. [Shot 2, 00:03.3–00:06.6] Cut to an elegant close-up of <Subject 2> wearing <Product 2>. She turns her chin a few degrees, catching an architectural reflection on the chrome frame. Same atrium, consistent light. [Shot 3, 00:06.6–00:10] Cut to a symmetrical medium-wide composition of both subjects standing side by side under the skylight: <Subject 1> wearing the red-band jacket and <Subject 2> wearing the chrome glasses. Both face the camera, with minimal natural movement. End on a clean editorial pose. Preserve identities and products across all three cuts, natural hands, no subject merging.

overall_soundscape: A single slow sneaker footstep in Shot 1, spacious concrete ambience throughout, slight jacket rustle, quiet transition whooshes at the two cuts.

non_diegetic_music: A minimal atmospheric electronic fashion cue with a low pulse, three soft metallic accents and a smooth ending.

Setup Configs

Three-shot capsule collection mini-commercial Duration: 10 seconds; choose 243 frames / 24 fps. Suggested FIRST test output: about 1 megapixels, preserve reference aspect ratio via the official template if possible. Ref_image_size: match. Steps: 8 | guidance: 1.0 | sigma shifts: video 12.0 / audio 3.0 | VSA: 10%

Connect ordered inputs into ref_images on MiniMaxH3ReferenceToVideo: <Picture 1> -> Picture_1.png -> first model identity <Picture 2> -> Picture_2.png -> second model identity <Picture 3> -> Picture_3.png -> shared featured puffer jacket <Picture 4> -> Picture_4.png -> featured chrome sunglasses <Picture 5> -> Picture_5.png -> shared location

Do not connect these images to the First Last Frame to Video node.

Actual render quality and VRAM requirements will vary.

Output

Generations Timings

Conclusion

These experiments show why reference-based video generation is worth testing beyond simple image-to-video prompts. FastH3 Omni gives us a way to combine several visual references and direct how they should appear in the final scene, but the results still depend on the prompt, reference images, and generation settings.

The most useful part of this process was comparing the outputs across different tasks instead of relying on one successful clip. If you're experimenting with AI video in ComfyUI, the prompts and configurations here should give you a practical starting point for your own tests. I'd be interested to see how the results change with different reference images, longer sequences, and hardware configurations.

FastVideo FastH3 Omni on HexGrid.cloud

Run FastVideo FastH3 Omni on HexGrid in one click — no CUDA setup, no dependency hunting, no environment debugging.

The template comes preconfigured with ComfyUI, the required runtime, and the FastVideo FastH3 Omni setup so you can launch a GPU workspace and start generating immediately.

No manual CUDA installation. No PyTorch version matching. No broken requirements. No hours spent wiring the environment together.

Just choose the FastVideo FastH3 Omni template, start the instance, and open ComfyUI.

Pricing starts from $0.58/hour*, so you can spin up a GPU only when you need it and shut it down when you’re done.*

Gemma-4 31B + vLLM + RTX 6000 PRO : 1168 tokens/sec and still asking for more...

Nick K. — Tue, 30 Jun 2026 12:35:13 +0000

We pushed Gemma-4 31B to 24 concurrent requests on a single RTX 6000 PRO Blackwell. The queue never filled. ~1.17k tokens/sec, and it still had headroom.

Most LLM "benchmarks" show you one request at a time. That tells you almost nothing about production.

So we ran Gemma-4 31B (FP8) on vLLM under a real ShareGPT workload, ramping concurrency 12 → 16 → 20 → 24, and watched what actually happens.

The numbers that mattered:

→ Peak throughput: ~1,168 tokens/sec total (~548 tok/s output)

→ Median time-to-first-token: ~0.7s — snappy even under load

→ Queue depth: averaged 0.41, peaked at just 3 while 14–21 requests ran concurrently

→ Server stayed unsaturated across the entire sweep

The one thing to watch:

Tail TTFT.
Median first-token stays fast, but p99 climbs to ~19s at the heaviest concurrency. That's the first metric to flex as you push higher — not throughput, not the queue.

Setup:

1× RTX 6000 PRO Blackwell (96GB)
Gemma-4 31B-it, FP8 checkpoint
vLLM 0.20 — prefix caching + chunked prefill on
ShareGPT workload, 1024 max output tokens, streaming ON
Max model length (context) : 4096

Verdict:

A single Blackwell card runs a 31B model at 24-way concurrency without breaking a sweat. The high end-to-end latency is just long generations, not queuing — and there's clearly room to climb past 24.

Token Throughput chart:

E2E Latency Chart

Full writeup — configs, charts, and per-concurrency breakdown — in the comments. ↓

Why NVIDIA Nemotron 3 Nano matters for private open-source inference. And an easy way to deploy it privately.

Nick K. — Sun, 21 Jun 2026 04:58:00 +0000

Nemotron 3 Nano is a 30B-class open model from NVIDIA built for efficient reasoning, coding, chat, agentic workflows, and long-context applications. It uses a hybrid Mixture-of-Experts architecture, activating only a small fraction of its total parameters per token, which makes it especially attractive for teams that want strong reasoning performance without the operational cost of serving much larger dense models.

Why Nemotron 3 Nano matters

NVIDIA describes Nemotron as a family of open models with open weights, training data, and recipes designed for building specialized AI agents.[¹]

Nemotron 3 Nano is the compact, efficiency-focused model in the Nemotron 3 family. It is designed for teams that need a practical balance of reasoning quality, inference throughput, long-context support, and deployment cost.

The model is especially interesting because it combines three important ideas:

Reasoning and non-reasoning in one model

Nemotron 3 Nano can generate reasoning traces before producing its final answer, and this behavior can be controlled through the chat template.[²]
Hybrid Mamba-Transformer MoE architecture

NVIDIA’s model card describes the model as a hybrid architecture with Mamba-2, attention, and Mixture-of-Experts layers. It has 30B-class total parameters while activating only around 3B-class parameters per token.[²]
Long-context support up to 1M tokens

NVIDIA’s Nemotron 3 research page highlights long-context support up to 1M tokens, making the model relevant for large documents, retrieval-heavy workflows, agent memory, codebases, logs, and enterprise knowledge tasks.[³]

This is what makes Nemotron 3 Nano compelling: it is not just a smaller model. It is a model designed to make advanced reasoning more practical to deploy.

Press enter or click to view image in full size

Model Snapshot

Press enter or click to view image in full size

Benchmark highlights

NVIDIA’s published benchmark table compares Nemotron 3 Nano 30B-A3B-BF16 against Qwen3–30B-A3B-Thinking-2507 and GPT-OSS-20B across reasoning, coding, agentic, instruction-following, long-context, and multilingual tasks.

Reasoning and coding performance

Press enter or click to view image in full size

NVIDIA’s benchmark data shows Nemotron 3 Nano performing strongly across math, coding, and formal reasoning tasks, with especially large gains on MiniF2F compared with the listed baselines.

What this means for developers

Benchmarks are useful, but the practical question is simpler:

What can I build with this model?

Nemotron 3 Nano is a strong fit for workloads where teams need a model that can reason, follow instructions, write code, use tools, and handle longer context — without deploying a giant dense model.

Good use cases include:

AI coding assistants
Agentic workflows
RAG over large document collections
Private enterprise chat
Technical Q&A
Math-heavy and logic-heavy workflows
Long-context document analysis
Workflow automation agents
Internal developer tools
Customer support assistants with private data

If your application needs more than lightweight chat but you still care about inference cost, Nemotron 3 Nano is a practical model to evaluate.

Agentic and tool-use performance

Nemotron 3 Nano is competitive across agentic benchmarks, with strong results on SWE-Bench, Terminal Bench, TauBench average, and BFCL v4.

Built for agentic AI

The most important shift in AI applications is that models are no longer just answering questions. They are planning, calling tools, reading long context, generating code, debugging workflows, and acting as sub-agents inside larger systems.

NVIDIA positions Nemotron models as open models for building specialized AI agents.[¹]

That makes Nemotron 3 Nano especially relevant for teams building systems where the model needs to:

break down tasks,
reason through multi-step problems,
work with tools,
operate inside workflows,
analyze long context,
and produce structured outputs for downstream systems.

For many teams, this is the gap between a chatbot demo and a production AI application.

Instruction-following and chat quality

Result: Nemotron 3 Nano shows strong instruction-following and chat performance, including a 67.7 Arena-Hard-V2 average in NVIDIA’s published model-card results.

Long context: built for large inputs

Modern AI applications often need more than a short prompt.

A production model may need to reason over:

long contracts,
research papers,
support histories,
code repositories,
logs,
product documentation,
multi-turn agent memory,
or large retrieved context windows.

NVIDIA’s Nemotron 3 research page states that Nemotron 3 models support context length up to 1M tokens, and highlights Nemotron 3 Nano’s performance on long-context RULER evaluations.[³]

The Hugging Face model card reports the following RULER scores:

Result: NVIDIA’s model-card results show Nemotron 3 Nano maintaining strong RULER scores from 256k to 1M context length

Throughput advantage

Performance is not only about benchmark accuracy. For production inference, throughput matters.

NVIDIA’s Nemotron 3 research page states that on an 8K input / 16K output setting with a single H200, Nemotron 3 Nano provides 3.3x higher inference throughput than Qwen3–30B-A3B and 2.2x higher throughput than GPT-OSS-20B.[³]

That matters because throughput directly affects:

user experience,
queueing delay,
infrastructure cost,
concurrency,
and how many requests your deployment can serve.

For production AI teams, a model that is both capable and efficient can be more valuable than a larger model that is harder to serve.

Why deploy Nemotron 3 Nano on HexGrid.cloud?

You can download open model weights yourself. You can run vLLM or SGLang yourself. You can provision GPUs yourself.

But then you own the entire infrastructure path.

That includes:

GPU provisioning,
runtime configuration,
model loading,
object storage,
HTTPS endpoint setup,
authentication,
logs,
usage tracking,
scaling,
billing visibility,
and production reliability.

HexGrid.cloud removes that operational overhead.

With HexGrid.cloud, Nemotron 3 Nano can be deployed as a private production endpoint on dedicated GPU infrastructure. Your app calls the model through the same OpenAI-compatible interface your existing tools already understand.

A typical request looks like this:

curl https://api.hexgrid.cloud/v1/chat/completions \ -H "Authorization: Bearer $HEXGRID_API_KEY" \ -H "Content-Type: application/json" \ -d '{ "model": "nemotron-3-nano-30b-a3b", "messages": [ { "role": "user", "content": "Explain why hybrid MoE models are useful for efficient reasoning." } ] }'

Same interface.
Private runtime.
Dedicated GPU deployment.

Final thoughts

Nemotron 3 Nano 30B-A3B is one of the most interesting open models for teams that want strong reasoning without giving up deployment efficiency.

It brings together:

30B-class model capacity,
roughly 3B-class active-parameter efficiency,
reasoning and non-reasoning modes,
strong coding and math benchmark results,
competitive agentic performance,
long-context support up to 1M tokens,
and a practical path for production inference.

If you are building agents, RAG systems, coding assistants, enterprise chat, or private AI applications, Nemotron 3 Nano is ready to test on HexGrid.cloud today.

Footnotes

[¹]: NVIDIA Developer — NVIDIA Nemotron: https://developer.nvidia.com/topics/ai/nemotron

[²]: Hugging Face — NVIDIA-Nemotron-3-Nano-30B-A3B-BF16 model card: https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16

[³]: NVIDIA Research — Nemotron 3 Family of Models: https://research.nvidia.com/labs/nemotron/Nemotron-3/

[⁴]: NVIDIA NIM model card — nemotron-3-nano-30b-a3b: https://build.nvidia.com/nvidia/nemotron-3-nano-30b-a3b/modelcard

[⁵]: arXiv — “Nemotron 3 Nano: Open, Efficient Mixture-of-Experts Hybrid Mamba-Transformer Model for Agentic Reasoning”: https://arxiv.org/abs/2512.20848

[⁶]: Hugging Face — NVIDIA-Nemotron-3-Nano-30B-A3B-FP8 model card: https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-FP8

[⁷]: HexGrid.cloud homepage: https://hexgrid.cloud/