Budget on-prem-ish batch open-model jobs

Scores as of 2026-10-19 (8 days old).

Depot Pick

SaladCloud · 7.17/10 for this job (curated estimate, high confidence)

Batch/cost focus. Confidence: medium — Salad price floor moves; pin /pricing.

Runner-up

Vast.ai · 7.41/10

Scores higher than the pick on: Capability fit, Price / performance, Developer & agent DX, Lock-in risk (higher = worse), Maturity / track record.

When not to use the pick

Need predictable enterprise GPU (CoreWeave/Nebius) or TEE (Chutes).

Lock-in / migration

Runner-up Vast.ai has lower stored lock-in risk (2.5 vs 3.5, 0–10, higher = more lock-in). Self-host fallback listed for this job: SmolLM2-1.7B-Instruct, MLX, llamafile.

Sourced price (usd_per_gpu_hour)

Pick's lowest sourced SKU: RTX 3090 24GB (Lowest priority, batch) $0.09 (usd_per_gpu_hour) — source, as of 2026-10-19

Cheapest matching SKU: SaladCloud RTX 3090 24GB (Lowest priority, batch) $0.09 — source, as of 2026-10-19. The pick is the cheapest.

Price history: first seen 2026-10-19 · last changed no change recorded · history JSON

Free option

• exo (oss) — source, as of 2026-10-25
• llamafile (oss) — source, as of 2026-10-25
• MLX (oss) — source, as of 2026-10-25

Failure modes to watch

• Runaway spend: Set a budget and a tool-call cap per task. Alert when spend jumps. Stop the run at the cap instead of summarizing a partial trace as success.
• Partial tool success: Require a per-item status and a failure count. Success for the task requires the failed count to be zero or explicitly handed to a retry queue.
• Context overflow: If the input exceeds the window, say so and retrieve a narrower slice. Do not answer from a truncated head as if it were the whole source.
• Lost in the middle: Ask the agent to name the passage it used. Spot-check questions whose only support sits in the middle of the context, not only at the edges.
• Non-determinism: Run the task more than once and compare the end state, not a single transcript. Record the trial count next to any success rate you cite.
• Schema drift: Validate every tool call and every loaded table against the current schema. Quarantine mismatches instead of coercing them.

Reported outcomes (submitted reports only)

This job: loading… JSON · reports never change scores · seed/test records excluded

Depot Pick (SaladCloud) on this job: loading… JSON · reports never change scores · seed/test records excluded

Same-job, same-unit comparisons: SaladCloud vs Chutes

Cite: Agent Provider Depot, Depot Pick for on-device-edge-batch: salad, scores as of 2026-10-19. Agents: GET /api/decide?jtbd=on-device-edge-batch