Which jobs your hardware can take
The Supernet checks two separate limits before sending a job to a Fast Worker. GPU VRAM is the memory on the graphics card. Host RAM is the system memory the worker actually has, meaning what Docker, WSL2 or the hosting container gives it, not what is installed in the PC. The two are never added together. A 96 GB card with 38 GiB of usable RAM still misses the jobs that need 47 GiB, and a PC with 128 GB installed still misses them if WSL2 only gives Docker 63 GB.
Hardware isn't the only gate. A job also needs:
- A current worker version. New modes need a minimum Sogni Comfy Worker release. Older workers are never offered them.
- Subscription-earnings opt-in for Unlimited-plan jobs. These go only to workers whose account has opted in and accepted the current terms.
- The Fast network. Linux and Windows NVIDIA workers join Fast by default. Jobs queued on Relaxed go to Relaxed workers.
- The model files on disk. A worker advertises only the workflows it has downloaded and can run.
The 12 GB support below is current as of October 31, 2026, with Comfy Worker 1.0.249. Other per-model thresholds retain their September 2026 baseline unless noted. Sogni tunes them from production completion and out-of-memory data, and they change. See Thresholds change.
#Contents
- Choosing a host: Internet and RAM: minimum connection targets and recommended capacity for video workers.
- How RAM and VRAM are measured: read this first if your numbers look lower than you expect.
- 12 GB NVIDIA GPUs: supported hardware variants, tested cards and available jobs.
- Quick reference by GPU
- Quick reference by host RAM
- Per-model detail: resolution, clip length, reference video and two-stage rules. Read this when a particular job type never reaches you.
- Get more jobs: a setup checklist.
- Thresholds change
#Choosing a host: Internet and RAM
When choosing a machine or rental GPU, use these connection targets:
| Connection | Practical minimum | Recommended |
|---|---|---|
| Download | 100 Mbps | 600 Mbps or faster |
| Upload | 100 Mbps | 250โ500 Mbps or faster |
These are sizing targets for efficient operation, not connection-speed eligibility checks. Download speed affects model and job-input transfers; upload speed affects how quickly completed results reach the artist. A 50 Mbps upload connection is a slower fallback, below the practical minimum we recommend for a new rental.
Use sustained speeds available to the worker, not just the host's advertised link speed. If several GPUs share a connection, allow enough bandwidth for their combined transfers. Mbps means megabits per second; 100 Mbps is about 12.5 MB per second before overhead.
For video model workers, we recommend 60 GB or more of system RAM available to each worker. Allow additional RAM for the operating system and other applications. Check the Docker, WSL2 or hosting-container allocation before renting or installing. Individual workflows have different minimums, listed below; the two-stage Reference-to-Video 1080p/2K workflow requires 60 GiB of usable RAM (about 64.4 GB), which is more than a 60 GB allocation.
#How RAM and VRAM are measured
- RAM is measured in GiB (1 GiB = 1,073,741,824 bytes), not the marketing GB. 47 GiB is about 50.5 GB.
- The container's limit counts, not the host's total. The worker reports the memory limit of its container (its cgroup), plus current-connection health readings of effective and total memory. The Supernet uses the smallest of those values.
- WSL2 gives Docker only half of Windows RAM by default. On a 64 GB PC that is about 31 GiB, and a
memory=48GBsetting shows up as roughly 46 to 47 GiB, which can fall just short of the 47 GiB video floor. Usememory=50GBor more if the PC can spare it. See Get more jobs. - VRAM is the figure the worker reports for its GPU. Some rules use 23 GB rather than 24 GB, and 30 GB rather than 32 GB, because cards often report slightly less than their nominal size.
- 12 GB Comfy workers: Comfy Worker 1.0.249 or later on CUDA 13 can take the qualified jobs listed below with 12 GB of VRAM. The Fast network still requires 30 GiB of reported usable host RAM. Existing classic-SD-only Comfy workers can qualify from 8 GB of VRAM; that exception does not enable other models. The worker checks the RAM floor before it connects and waits instead of looping when RAM is insufficient.
- GPU generation and driver (Comfy Worker): an NVIDIA RTX 30-series (Ampere) or newer GPU with driver R580 or newer. The installer refuses older GPUs and drivers and says which one to update.
#12 GB NVIDIA GPUs
On an x86-64 Windows or Linux host, the CUDA 13 Comfy worker accepts GPUs reporting compute capability 8.0 or newer, an R580-or-newer NVIDIA driver, and enough VRAM and host RAM for the selected job. These are the manufacturer-listed 12 GB variants meeting that architecture requirement as of October 31, 2026:
| Type | 12 GB GPU variants |
|---|---|
| GeForce desktop | RTX 3060 12 GB, RTX 3080 12 GB, RTX 3080 Ti, RTX 4070, RTX 4070 SUPER, RTX 4070 Ti, RTX 5070 |
| GeForce laptop | RTX 4080 Laptop GPU, RTX 5070 Laptop GPU 12 GB, RTX 5070 Ti Laptop GPU |
| Professional desktop | RTX A2000 12 GB |
| Professional laptop | RTX A3000 12 GB Laptop GPU, RTX 3500 Ada Generation Laptop GPU, RTX 4000 Ada Generation Laptop GPU, RTX PRO 3000 Blackwell Generation Laptop GPU |
| Embedded GPU on an x86-64 host | RTX 3500 Ada Generation Embedded GPU |
Physical qualification: the 12 GB workload tests ran on an RTX 3060 and an RTX 5070. The other entries meet the architecture and manufacturer memory specifications; they have not each received the same physical qualification. The installer checks the actual GPU capability and driver rather than relying on its product name. Laptop and embedded power limits can affect throughput.
Check the exact memory variant. The RTX 3060 also has an 8 GB desktop version, the RTX 3080 has a 10 GB version, and the RTX 5070 Laptop GPU has both 8 GB and 12 GB versions. A desktop RTX 4070 Ti SUPER has 16 GB and belongs to a different class. The RTX 2060 12 GB does not qualify for the current CUDA 13 worker: its Turing architecture is below the worker's compute-capability floor.
Manufacturer references: GeForce desktop specifications, GeForce laptop specifications, professional laptop specifications, RTX A2000 12 GB specifications, RTX A3000 12 GB laptop family, embedded specifications, and CUDA compute capabilities.
#Jobs available on 12 GB workers
- Stable Diffusion 1.5 and SDXL through the classic model collections.
- FLUX 3 Image Klein INT8 ConvRot: image generation and editing, including compatible LoRAs. Jobs above a 2048 ร 2048 pixel area retain the existing 32 GB-class routing requirement.
- Qwen3-TTS CustomVoice: speech generation with preset voices.
- BiRefNet: background removal, when its optional model pack is selected.
- SAM 3: object segmentation and masks, when its optional model pack is selected.
The current worker advertises only eligible workflows with the required files and capabilities. Voice cloning, voice design, ACE-Step music and heavier image/video models retain their existing VRAM floors; experimental engine runs do not add them to this supported list. See the 1.0.249 release notes.
#Quick reference by GPU
Each column assumes the machine also meets the RAM floor in the next table.
| GPU class | Examples | What it can take |
|---|---|---|
| 12 GB | RTX 3060 12 GB, 4070, 5070; all variants above | Classic SD1.5/SDXL, FLUX 3 Image Klein, Qwen3-TTS CustomVoice, optional BiRefNet and SAM 3 on CUDA 13 worker 1.0.249+. Requires 30 GiB usable host RAM. |
| 16 GB | RTX 4080, 4070 Ti Super, 5080 | Image models such as Flux, Chroma, Z-Image, Krea 2 and One Obsession (Z-Image and Krea 2 up to about 2.5 MP); ACE-Step 1.5 music; upscales, which are routed to smaller cards first. No video. |
| 20 GB | RTX 3080 20 GB class | Everything above, plus ACE-Step 1.5 XL songs up to 8 minutes |
| 24 GB | RTX 4090, 3090 | Everything above, plus Z-Image and Krea 2 above 2.5 MP; Qwen Image Edit 2511; MiniMax Music 3; Qwen3 voice models; full-length ACE-Step XL; WAN 2.2 under 700p at 8 steps or fewer; LTX up to about 2.2 MP (1080p); MiniMax H3 FastH3 Turbo in every mode, including custom LoRAs and audio-guided modes; FastH3 two-stage 768p/1080p/2K for clips under 175 frames; FlashVSR |
| 32 GB | RTX 5090 | Everything above, plus H3 Standard, Balanced and LightX2V Turbo; Reference-to-Video from images; two-stage Reference-to-Video at 768p, 1080p and 2K with reference videos, at any length; LightX2V Turbo Reference-to-Video with one reference video, inside the RTX 5090 envelope; LTX up to about 4 MP (1440p); FastH3 two-stage 768p and 1080p at any length; LTX 10Eros; WAN at 700p+ with 8 steps or fewer, or under 700p with more than 8 steps; Pixal3D image-to-3D |
| 48 GB | RTX 6000 Ada, L40S, A6000 | Everything above, plus FastH3 two-stage 2K clips of 175+ frames; one-stage Reference-to-Video with a reference video, under 243 frames and up to 0.9 MP; Standard Reference clips of 342+ frames; LTX 4K; WAN 700p+ with more than 8 steps |
| 96 GB | RTX PRO 6000 | Every job type, including one-stage Reference-to-Video with a reference video at 243+ frames or above 0.9 MP, long Balanced Reference at the full 768p canvas, and LTX-2.5 4K clips longer than 5 seconds. 80 GB data-center cards qualify for the last one too. |
The Supernet treats cards reporting 40 GB or more as scarce capacity. See Why 48 and 96 GB cards get certain jobs first.
#Quick reference by host RAM
Usable RAM, as measured above. "Small GPU" means a card reporting 32 GB or less. Larger cards need less RAM for most MiniMax H3 and LTX work.
| Usable RAM | Small GPU (24/32 GB) | Large GPU (48/96 GB) |
|---|---|---|
| Under 30 GiB | Not admitted to the Fast network | Not admitted |
| 30 to 38.99 GiB | Images, music, voice, upscales, WAN, Pixal3D, RelayVSR, and FlashVSR on cards reporting 31 GB or more. No MiniMax H3 and no LTX. | Everything except FastH3 two-stage 2K and the Reference-to-Video 1080p/2K cascade |
| 39 to 46.99 GiB | Adds H3 Standard and Balanced: text, image, first/last-frame and Reference, plus two-stage Reference at 768p | Same as above |
| 47 to 59.99 GiB | Adds FastH3 Turbo (all modes), LightX2V Turbo (including Reference), FastH3 two-stage 768p, 1080p and 2K, LTX-2.3/2.5, and FlashVSR on 24 GB cards | Adds FastH3 two-stage 2K |
| 60 GiB or more | Adds two-stage Reference-to-Video at 1080p and 2K | Adds two-stage Reference-to-Video at 1080p and 2K |
#Per-model detail
#MiniMax H3: minimums by tier
| Tier | Minimum VRAM | RAM on a small GPU | RAM on a large GPU | Worker version |
|---|---|---|---|---|
| FastH3 Turbo (text, image, first/last frame, custom LoRAs) | 23 GB | 47 GiB | 30 GiB | 1.0.193 |
| FastH3 audio-guided (image + audio, first/last + audio, audio only) | 23 GB | 47 GiB | 30 GiB | 1.0.217 |
| LightX2V Turbo (text, image, first/last frame) | 32 GB | 47 GiB | 30 GiB | 1.0.208 |
| LightX2V Turbo Reference-to-Video | 32 GB | 47 GiB | 30 GiB | 1.0.171 |
| Standard and Balanced (all modes, including Reference) | 32 GB | 39 GiB | 30 GiB | 1.0.155 and up, by mode |
Attaching a custom LoRA to an H3 job needs worker 1.0.176 or later. A Reference job with a reference video needs 1.0.177 or later.
Windows (WSL2) exception: LightX2V Turbo first/last-frame jobs go only to WSL2 workers with more than 40 GB of VRAM, so an RTX 5090 on Windows does not receive them. On native Linux, the 32 GB floor applies.
#Two-stage output (768p, 1080p and 2K)
A two-stage job renders at a half-size canvas and delivers at 2ร. The canvas sets the class: a short edge up to 384 px and a long edge up to 672 px delivers 768p, a short edge up to 544 px delivers 1080p, and anything larger delivers 2K.
| Two-stage job | VRAM | Host RAM | Worker version and files |
|---|---|---|---|
| FastH3, 768p/1080p class | 23 GB under 175 frames; 30 GB (the 32 GB class) at 175 frames or more | 47 GiB on a small GPU, 30 on a large one | 1.0.217 and the H3 latent upscaler |
| FastH3, 2K class | 23 GB under 175 frames; 40 GB at 175 frames or more | 47 GiB on every GPU | Same |
| Reference (Standard/Balanced), 768p class | 32 GB at any length, with or without reference video | 39 GiB on a small GPU, 30 on a large one | 1.0.218 and the latent upscaler |
| Reference (Standard/Balanced), 1080p and 2K | 32 GB at any length, with or without reference video | 60 GiB on every GPU | 1.0.220, the latent upscaler and the FastH3 checkpoint (CUDA 13 image) |
Long FastH3 two-stage clips (175 frames or more) need 40 GB at the 2K class and 30 GB (the 32 GB class) at 768p and 1080p. An RTX 5090 peaked at about 28.9 GB (768p) and 29.3 GB (1080p) on a 15-second clip, because those classes render their first pass at half size.
#Reference-to-Video with a reference video
One-stage Reference-to-Video, whether Standard, Balanced or LightX2V Turbo, uses much more memory when an artist attaches a reference video:
- Under 243 frames, at 0.9 MP or less: more than 40 GB (the 48 GB class).
- At 243 frames or more, or above 0.9 MP (for example 1344ร768): more than 48 GB (the 96 GB class).
- RTX 5090 envelope for LightX2V Turbo only: a card named RTX 5090 with 32 GB takes the job when it has exactly one reference video, at most one reference image, no separate audio reference and no custom LoRA, and either:
- no reference image, up to 1120ร480 pixels (537,600), up to 158 frames, with pixels ร frames up to 1120 ร 480 ร 124; or
- up to 960ร544 pixels (522,240) with the long side at most 960, up to 362 frames.
Two-stage Reference jobs don't follow these rules. They render the reference-conditioned pass at a quarter of the final size, so a 32 GB card takes them (see the table above).
#Long Reference-to-Video clips (image references)
One-stage Standard and Balanced Reference clips of 342 frames or more need 40 GB. Long Balanced clips above 0.9 MP need 49 GB, which means the 96 GB class. LightX2V Turbo Reference is exempt. Two-stage Reference IDs are exempt.
#LTX-2.3 and LTX-2.5
- Baseline: 23 GB of VRAM. Host RAM must be 47 GiB on 24/32 GB cards and 30 GiB on larger ones. The rule is checked at model load, so it applies to every LTX job.
- Resolution: under 30 GB, up to about 2.2 MP (1080p class). Under 40 GB, up to about 4.0 MP (1440p class). At 40 GB and above, full 4K. Size is judged on the grid-aligned size the worker actually renders.
- RTX 3090/4090 exception: a desktop 3090 or 4090 (not a laptop GPU) can take short 4K LTX-2.3 clips of up to 72 frames.
- Long clips (122+ frames): above 320 frame-megapixels (width ร height ร frames รท 1,000,000) needs 30 GB, and above 600 needs 40 GB. 1080p-class clips (2.0 MP or more) longer than 360 frames need 30 GB, and 48 GB+ cards are offered them first.
- LTX-2.5 at 4K longer than 121 frames: 80 GB, which means the 96 GB class.
- 4K longer than 361 frames: no worker class currently qualifies.
- LTX-2.3 10Eros: 30 GB for every job.
#WAN 2.2
| Resolution (short side) | 8 steps or fewer | More than 8 steps |
|---|---|---|
| Under 700 px | 23 GB | 32 GB |
| 700 px or more | 32 GB | 40 GB |
#Images, music, voice, upscaling and 3D
- Z-Image and Krea 2 family (including Identity Edit and Dark Beast): 16 GB, or 24 GB above 2.5 MP (roughly larger than 1536ร1536).
- Qwen Image Edit 2511: 24 GB. The worker does not offer it, or Z-Image BF16, on RTX 30-series cards (compute capability 8.6), because those cards return corrupt images for these models.
- FLUX 3 Image Klein: 12 GB on CUDA 13 worker 1.0.249+, subject to its existing job-size rules. Qwen3-TTS CustomVoice: 12 GB. Qwen3-TTS voice cloning and voice design: 24 GB.
- BiRefNet and SAM 3: available to qualified 12 GB workers with the corresponding optional model packs. Their download-pack floors alone do not override network admission or job requirements.
- ACE-Step 1.5 XL: 20 GB for songs up to 8 minutes, and 23 GB for longer ones. MiniMax Music 3: 24 GB.
- FlashVSR video upscaling: 23 GB. On cards reporting under 31 GB, FlashVSR keeps its caches in host RAM and needs 47 GiB. At 31 GB or more, the 30 GiB fleet floor applies.
- RelayVSR video upscaling: 23 GB, the CUDA 13 worker image and worker 1.0.263. The fleet RAM floor applies on every card.
- Image upscales run on any card. Cards with 23 GB or more let smaller cards take them first.
- Pixal3D image-to-3D: 30 GB, the CUDA 13 worker image and worker 1.0.208 (1.0.217 for multi-view). It is an optional model pack of about 15.5 GB, offered to eligible workers.
#Get more jobs
#1. Give the worker enough RAM
Windows (Docker Desktop on WSL2). Edit %USERPROFILE%\.wslconfig as the user who runs Docker Desktop, or answer y when the installer offers to update it. Then stop the worker, quit Docker Desktop, run wsl --shutdown, and start again.
| PC RAM | Suggested [wsl2] setting |
About what the worker sees | What it unlocks on a 24/32 GB card |
|---|---|---|---|
| 32 GB | Default only allows ~15 GiB | Below 30 GiB | Not enough for the Fast network. Use Linux or add RAM. |
| 64 GB | memory=50GB (48GB leaves more for Windows but can report just under 47 GiB) |
about 49 GiB | All H3 tiers, LTX and FlashVSR on 24/32 GB cards |
| 96 GB | memory=64GB |
about 62 GiB | Everything, including the Reference 1080p/2K cascade |
| 128 GB+ | memory=64GB or more |
62 GiB or more | Everything |
On a 64 GB PC, memory=50GB leaves about 14 GB for Windows, which is fine on a dedicated worker; leave more if you use the PC for other work. On larger PCs, leave at least 16 GB. Adding swap does not raise the usable-RAM figure the Supernet checks.
Linux, Docker, Nosana and Salad. The Supernet reads the container's memory limit. Remove or raise any --memory / mem_limit setting, and on hosting platforms choose a container memory allocation, not just a GPU. We recommend 60 GB or more of system RAM available to each video worker. H3 Turbo and LTX on a 24/32 GB card need at least 47 GiB usable RAM; two-stage Reference-to-Video at 1080p/2K needs 60 GiB. Check the reported usable RAM against the job's floor. A 48 GB allocation often reports 46.x GiB, just under the 47 GiB floor. Some hosts report much less RAM than their GPU class suggests. For example, 96 GB confidential VMs have reported about 38 to 39 GiB, which misses FastH3 2K and the Reference cascade.
#2. Keep the worker updated
Two-stage FastH3 needs 1.0.217, two-stage Reference 768p needs 1.0.218, and Reference 1080p/2K needs 1.0.220. Older workers are never offered these jobs. Keep persistent storage so a new model's files download once and stay.
#3. Opt in to subscription earnings
Unlimited-plan jobs go only to workers whose account opted in and accepted the current terms at dashboard.sogni.ai/subscription-earnings. Without that, the worker only receives Spark- and SOGNI-paid jobs.
#4. Don't narrow the workflow list without a reason
The worker advertises only the workflows it serves and has downloaded. Removing video workflows in Settings, or using a slimmer hosting preset, removes those jobs.
#Why 48 and 96 GB cards get certain jobs first
A card reporting 40 GB or more is first offered the queued jobs whose VRAM floor is above 32 GB, highest floor first. Examples are long two-stage FastH3, video-referenced Reference-to-Video, LTX 4K, long LTX clips and high-step WAN. This keeps work that only a large card can do from waiting behind work any RTX 4090 or 5090 can do. It is a preference, not a filter: when no such job is queued, a large card takes ordinary work. Payment priority is taken into account too, so a large card can reach past a higher-priority ordinary job to take a job only it can run.
#Check what the network sees
Open your worker at dashboard.sogni.ai/fast-workers. The worker view shows its version and GPU, and Worker Health charts VRAM and host memory for the current run. See Using the Workers Dashboard. On Windows, worker-start.bat also reports Windows RAM, Docker's usable WSL RAM and your .wslconfig setting. If a machine should qualify but a job type never arrives, compare its host memory with the RAM table above before restarting anything.
#Thresholds change
The 12 GB section reflects the October 31, 2026 release; other per-model thresholds retain the September 2026 baseline unless noted. Several are runtime settings Sogni can change without a release, including the long-clip frame counts and VRAM floors, the Balanced full-canvas rule, and the scarce-card steering. Floors are raised when a hardware class fails jobs, and lowered when production data shows it completes them. The worker's own catalog at api.sogni.ai/v1/worker/config/comfy lists the minimum VRAM for each model pack it downloads. Check this page again after each worker release.