LoRA Serving Options for SkillNet UI
Context
SkillNet needs to generate UI (A2TL-Web/JSON) using a trained LoRA. The goal is to serve the LoRA faster than a local RTX 4060 (~20-50 tok/s) without managing complex infrastructure.
Multi-LoRA is not needed. Only a single hosted LoRA is required.
Verified Options
| Option | Hosts Custom LoRA | Base Model | Approximate Cost | Approximate Speed | Ease |
|---|---|---|---|---|---|
| Groq LoRA | Yes, Enterprise-only | Officially only Llama 3.1 8B |
No public pricing | ~560 tok/s (base 8B) | Medium |
| Together AI | Yes, dedicated endpoint | Wide (Llama, Qwen, Mistral, etc.) | H100: $5.49/h per minute online | Not officially published | Easy |
| Fireworks AI | Yes, on-demand only | Wide (Llama, Qwen, DeepSeek, etc.) | H100: $7/h per minute online | ~90 tok/s on 8B H100 | Easy |
| Modal | Yes, with vLLM self-host | Any model supported by vLLM | H100: $3.95/h, L4: $0.80/h | Depends on GPU and config | Hard |
| Cerebras | Yes, Enterprise + preview | Llama, Qwen3-Coder, GPT-OSS, etc. | No public pricing | ~1,000-1,700 tok/s | Very hard |
Analysis by Option
Groq
- Clear process: upload a ZIP with
adapter_model.safetensors+adapter_config.json, register as fine-tuned model, call via API. - Hard limitation: only
Llama 3.1 8Baccording to current docs. The 70B mentioned in a 2025 announcement no longer appears. - Cold start proportional to rank (8, 16, 32, 64).
- Verdict: good if Enterprise access is granted and 8B is enough. Not self-service.
Together AI
- Trains and serves LoRAs. Fine-tuning from $0.48/1M tokens (≤16B).
- Serving requires a dedicated endpoint.
- Billed per minute of GPU uptime.
- Verdict: good for iteration, but the endpoint must be shut down to avoid continuous cost.
Fireworks AI
- Similar to Together. LoRA fine-tuning from $0.50/1M tokens (≤16B).
- Custom LoRAs only on on-demand deployments.
- Live merge offers the same performance as the base model.
- Verdict: balanced option to start.
Modal
- Cheaper than Together/Fireworks (L4 at $0.80/h, H100 at $3.95/h).
- Requires setting up vLLM, volumes, autoscaling.
- Cold starts if
min_containers=0. - Verdict: better long-term cost, but more work.
Cerebras
- Multi-LoRA in private preview for dedicated endpoints.
- Process for a single LoRA is not publicly documented.
- Enterprise required. No public pricing.
- Verdict: overkill for a single LoRA. Only for massive production scale.
Realistic Monthly Costs (Fireworks H100)
| Usage | Monthly Cost |
|---|---|
| 2 h/day | ~$15-20 |
| 8 h/day | ~$170 |
| 24/7 | ~$500 |
Modal L4 24/7: ~$580/month, but much cheaper per hour.
Recommendation for SkillNet
- Train the LoRA on Fireworks AI or Together AI (self-service, cheap).
- Serve it on Fireworks on-demand with a small model (8B-9B) and autoscaling to 0 to avoid continuous cost.
- If usage becomes regular, evaluate Modal with L4 to reduce cost.
- Groq or Cerebras only if Enterprise access is obtained and extreme speed is justified.
Next Steps
- Define the SkillNet A2TL-Web/JSON format.
- Generate 100-200 training examples.
- Train a test LoRA on Fireworks/Together.
- Measure speed and quality against local RTX 4060.
Sources
- Groq LoRA docs: https://console.groq.com/docs/lora
- Together AI fine-tuning: https://docs.together.ai/docs/fine-tuning/overview
- Together AI dedicated endpoints pricing: https://docs.together.ai/docs/dedicated-endpoints/pricing
- Fireworks fine-tuning: https://docs.fireworks.ai/fine-tuning/fine-tuning-models
- Fireworks pricing: https://fireworks.ai/pricing
- Modal pricing: https://modal.com/pricing
- Cerebras Multi-LoRA blog: https://www.cerebras.ai/blog/introducing-multi-lora-on-cerebras-inference
- Cerebras dedicated endpoints: https://inference-docs.cerebras.ai/dedicated/overview