Close-up of GPU silicon under cool studio light
Models

Choosing a model from the catalog

Pick by job shape and hardware, not by whichever name is trending this week.

Scalattice serves a fixed catalog. Developers call model IDs on the API. Providers download those same weights and earn when jobs land. That shared list is intentional: every machine that enables a model can take the same work.

Live rates and demand labels sit in Scalattice Cloud and on pricing. This post is about fit. For why the August 2026 families were added, see the catalog post.

The lineup

Llama 3.3 70B

Flagship general chat and long-context work. Highest quality in the catalog for open weights at this size. Needs serious VRAM on the provider side. Use it when answers matter more than pennies.

Qwen3 32B

Strong all-rounder under the 70B. Good default when you want quality without flagship cost or hardware. Providers with mid-high VRAM cards often land here.

Gemma 4

Google’s open mid-size: gemma-4-26b-a4b (MoE) and gemma-4-31b (dense). Different personality than the Qwen line. Useful if you want a second mid-tier option for A/B or vendor diversity.

Qwen3 14B

Sweet spot for many products: solid quality, lower latency and price than 32B, fits more GPUs. Default recommendation when you are not sure yet. Bake off against qwen-3.5-9b if 14B is only “fine.”

Qwen3 8B

Fast, cheap, and widely hostable. Classification, short chat, tool loops, and prototypes. Do not expect 70B nuance.

DeepSeek R1 7B

Reasoning-oriented distill. Prefer it for multi-step problems, math-ish tasks, and prompts where “think then answer” helps. Overkill for simple completions. gpt-oss-20b is the other reasoning stack if you want a different voice.

Qwen3 Coder 30B A3B

Code generation, edits, and repo-shaped prompts. 256K context is the point. Keep general chat on a general model; send coding traffic here.

Llama 4 Scout

The long-context specialist (10M). Use it when the job is a huge document or needle-in-haystack, not as a default chat model. Needs a fat GPU.

For developers

  • Start on 8B or 14B while you measure quality on your eval set.
  • Move up only when you can show a failure the smaller model keeps repeating.
  • Use the coder for code paths; use R1 or gpt-oss when chain-of-thought style work wins.
  • Pin model IDs in config. Catalog IDs are stable product names (for example qwen-3-14b).

For providers

  • Enable models your VRAM and disk can actually hold. Partial downloads stall readiness.
  • Demand labels are directional, not a guarantee of jobs. Capacity and schedules matter too.
  • Smaller models fit more cards and turn over faster. Large models pay more per token but need quieter, fatter GPUs.

A simple rule

If the task is short and frequent, go small. If a wrong answer is expensive, go large. If the job is code or reasoning, use the specialist. Everything else is tuning.