Model Selection: Which Model for Which Use Case?
The noris “LLM as a Service” platform offers several models of varying size and strength. The right choice depends on the use case: small models excel at speed and low cost, large ones at reasoning quality and context length. This decision matrix helps you get started quickly.
Decision Matrix
Section titled “Decision Matrix”| Use Case | Recommended Model | Why |
|---|---|---|
| Simple chat, classification | Gemma 4 31B | Fast, cost-effective, multimodal |
| Complex reasoning, math | GLM 5.3 | Controllable reasoning effort (reasoning_effort), chain-of-thought |
| Coding, repository-level | Qwen 3.8 27B | Terminal Bench 73.0%, SWE-bench Pro 61.7%, agentic coding |
| Highest coding and agent demands | GLM 5.3 | Terminal-Bench 2.1: 88.2%, DeepSWE: 66.9% (Experimental) |
| Long documents (1M+) | GLM 5.3 Flash or DeepSeek V4.1 Flash | 1M context, hybrid attention resp. CSA2 |
| RAG embedding | Harrier OSS 0.6B | Compact, fast, RAG-optimized |
| RAG reranking | BGE Reranker v2 M3 | Multilingual, precise |
| Multilingual RAG reranking | Jina Reranker v2 | 0.3B, high throughput |
Additional Factors
Section titled “Additional Factors”Beyond the primary use case, further criteria influence the choice:
- Context Length: How many tokens need to fit in a prompt? For long documents or entire repositories, prefer GLM 5.3 Flash or DeepSeek V4.1 Flash with up to 1M tokens of context. Smaller models are often sufficient if you’re already feeding short candidates via RAG.
- Cost: Larger models consume more tokens per request and cost more AI points. For bulk classification tasks, a compact model like Gemma 4 31B pays off.
- Latency: Larger models respond more slowly, especially with long contexts. For interactive chat applications with strict response-time requirements, prefer smaller models or enable Streaming.
- Tier: Prefer LTS versions for production environments. Experimental or short-cycle releases are suitable for testing and evaluation but shouldn’t be integrated directly into business-critical pipelines.
A complete overview of all available models, including context lengths and status, is available in the Models overview.
