LoRA serving & distillation hosting
Train adapters or distill smaller students, then serve many LoRAs efficiently without one-GPU-per-customer economics.
At a glance — Oct 2026
Updated 2026-10-27 · forward this page: /jtbds/lora-serving-distillation#briefScores as of 2026-09-14 (44 days old; Depot marks scores stale after 30 days). The pick is still Depot's current editorial pick; it has not been re-reviewed since. Prices below carry their own dates.
- Depot Pick
- Predibase · 7.88/10 for this job (curated estimate, medium confidence)
LoRAX-style multi-adapter serving is the differentiator. Confidence: medium — Predibase public token matrix uneven (see gaps).
- Runner-up
- Together AI · 8.13/10
Scores higher than the pick on: Capability fit, Price / performance, Latency / reliability, Developer & agent DX, Lock-in risk (higher = worse), Maturity / track record.
- When not to use the pick
- DIY Unsloth + your GPU cloud, Lightning Studios workflow, or Together/Fireworks finetune+serve already contracted.
- Lock-in / migration
- Runner-up Together AI has lower stored lock-in risk (3.5 vs 4.5, 0–10, higher = more lock-in).
- Sourced price (usd_per_million_train_tokens)
Pick: no sourced price in this job's unit (no_sourced_price).
Cheapest matching SKU: Fireworks AI Managed training LoRA SFT up to 16B $0.5 — source, as of 2026-10-03. Not the pick; not quality-adjusted.
Price history: first seen 2026-10-03 · last changed no change recorded · history JSON
- Free option
- No free resource lists this job. A cheap paid SKU is not a free option.
- Failure modes to watch
- No failure mode plainly applies to this job (no_failure_mode_plainly_applies).
Cite: Agent Provider Depot, Depot Pick for lora-serving-distillation: predibase, scores as of 2026-09-14. Agents: GET /api/decide?jtbd=lora-serving-distillation
All sourced prices, cheapest first
Predibase
LoRAX-style multi-adapter serving is the differentiator. Confidence: medium — Predibase public token matrix uneven (see gaps).
Pricing (directional): As of 2026-09-14: Active managed LoRA/QLoRA + RL fine-tune (Rubrik-era Predibase); LoRAX OSS multi-adapter serving. Public token rates still not a clean first-party matrix — secondary sources cite serverless from ~$0.20/M and A10G-class GPU…
Lock-in risk: higher = more lock-in (worse). Overall already inverts this dimension.
Alive. Prefer when you serve dozens of LoRA adapters on shared GPUs. Together/Fireworks remain simpler defaults for one-off SFT.
Quote for LoRA serving & distillation hosting
Depot Pick stays Predibase. Prices below are sourced SKU rows only (as of 2026-10-27). Input and output are not blended. GET /api/quote?jtbd=lora-serving-distillation
Only 1 candidate has a matching Train $/1M; nothing was compared, so pickIsPriceWinner is null.
Sourced price unavailable for the Depot Pick. No row in the pricing index.
3 of 5 candidates have no sourced price row. 4 are excluded from the Train $/1M comparison (train). min sourced training-token SKU among candidates; not chat inference input; not quality-adjusted
Runners-up
Together AI
Best catalog breadth for OSS models. Watch minimums and lifetime spend gates before you design around a gated…
Fireworks AI
Top pick for production open-model serving with real DX. Prefer over random GPU DIY when you want tokens not…
Unsloth
Depot OSS fine-tune default. Hosted LoRA → Predibase/Fireworks/Together; dead OpenPipe → do not resurrect…
Lightning AI
Use when the team already lives in Lightning/PyTorch notebooks. Together/Unsloth remain Depot finetune…
When NOT to use the top pick
DIY Unsloth + your GPU cloud, Lightning Studios workflow, or Together/Fireworks finetune+serve already contracted.
Full ranking (score table)
Ranked with this JTBD's weights. Scores are curated estimates unless marked otherwise. Agents: prefer /api/decide then this table for tie-breaks.
| # | Provider | Overall | Fit | $/perf | Latency | DX | As of |
|---|---|---|---|---|---|---|---|
| 1 | Unsloth | 8.5 | 8.8 | 9.2 | 7.5 | 9.0 | 2026-09-14 |
| 2 | Fireworks AI | 8.3 | 8.6 | 8.8 | 8.4 | 9.0 | 2026-09-14 |
| 3 | Together AI | 8.1 | 8.5 | 9.0 | 8.0 | 8.4 | 2026-09-14 |
| 4 | Lightning AI | 8.0 | 8.6 | 7.0 | 8.2 | 9.0 | 2026-09-14 |
| 5 | PredibaseDepot Pick | 7.9 | 8.4 | 8.0 | 7.8 | 8.2 | 2026-09-14 |
Score methodology for this job
LoRAX-style multi-adapter serving is the differentiator. Confidence: medium — Predibase public token matrix uneven (see gaps).
capabilityFit: 32%pricePerformance: 12%latencyReliability: 14%developerDx: 12%privacyPosture: 12%lockInRisk: 8%maturity: 10%