Model Catalog
Stage: Build · What it does: the curated set of open models available on the platform — the managed Bedrock models used platform-wide (Ask AI, judges, LLM Format, assistants, one-click deploy), the base models you can fine-tune with LoRA/QLoRA, and the frontier one-click catalog for full-weight serving.
Why it matters
The base model is the one choice you can't fix later: context length and multimodal ability are locked at base-model selection. Style and behavior are fixable via fine-tuning and post-training. Pick the base by the axis you can't change.
The platform has three distinct model surfaces — know which one you're picking from:
| Surface | What it is | Where you pick |
|---|---|---|
| Managed Bedrock models | 22 pre-integrated models used by platform features (Ask AI, LLM-as-judge, LLM Format, custom assistants, one-click deploy) | Ask AI, judge settings, LLM Format, Assistants, Ship → one-click |
| Fine-tune bases | HuggingFace base models (Llama 3, Mistral, Qwen, …) you adapt with LoRA/QLoRA — pick the base by name | Build → Fine-tune |
| One-click frontier catalog | Full-weight frontier models served on big GPUs, with context presets and honest per-hour costs | Ship → deploy from the catalog (see 17-one-click-catalog.md) |
Managed Bedrock models (the platform's 22)
Every model below is available in the Models settings tab, where an admin controls which models can be used for which purpose (see 16-model-governance.md). Metering is per-model: the credit multiplier is the real blended cost vs. the platform's baseline model, so enabling a model can never undercharge.
| Family | Models | Best for | Notes |
|---|---|---|---|
| DeepSeek | v3.2 (default), v3.1, V3, R1 (reasoning) | Default all-rounder; R1 for reasoning/math | v3.2 is the platform baseline; 131K context; available in 13 regions |
| Amazon Nova | Lite, Micro, Pro | Fast, cheap, long context (300K) | Micro is the lightest managed model |
| Llama (Meta) | 3 8B, 3.1 8B, 3.3 70B | Ecosystem baseline, tooling | 70B for higher quality on a budget |
| Mistral | 7B, Ministral 8B, Large 3 | Fast inference, strong per-parameter quality | Large 3 has 131K context |
| Qwen | Qwen3 32B, Qwen3 235B A22B | Strong multilingual, MoE efficiency | 32K context |
| GPT-OSS (OpenAI) | gpt-oss 20B, gpt-oss 120B | Open-weight OpenAI-class models | 131K context, MoE |
| Gemma (Google) | 3 4B, 12B, 27B | Lightweight, cheap capability | 131K context on 12B/27B |
| Claude (Anthropic) | 3 Haiku | Fast frontier-quality responses | 200K context |
| Titan (Amazon) | Embeddings v2 | Embeddings only | Powers the gateway's /v1/embeddings and /v1/rerank |
Honest scope: this is the full list of managed Bedrock models integrated today. If you need a model that isn't listed, the options are (a) fine-tune a HuggingFace base with LoRA/QLoRA and deploy it as your own endpoint, or (b) use the one-click frontier catalog for full-weight models. Ask in the console if you'd like a specific managed model added.
Fine-tune bases (LoRA/QLoRA)
Fine-tuning adapts a HuggingFace base model by name — the guided presets pre-fill sensible defaults (chat, instruction, SQL, support). You adapt a base; you never reproduce one. A 7–32B LoRA run is typically on the order of $1–3k of instance time, not the $0.5M–$25M the base cost to build.
Picking guidance for fine-tune bases:
| Need | Reach for |
|---|---|
| Default all-rounder | Qwen 3 class (strong multilingual, Apache-2.0) |
| Ecosystem baseline / tooling | Llama 3 / 3.1 / 3.3 |
| Long context (100K–1M) | Long-context variants of Llama 4 / Kimi K3 class (when available as a base) |
| Cheap capability per dollar | MoE: Qwen3-30B-A3B, GPT-OSS-20B |
| Multimodal input | Gemma 3, Qwen3-VL, Mistral 3 class |
| Code | Code-specialized variants of Qwen / DeepSeek |
| Reasoning on math/SQL/code | DeepSeek-R1 distills, Ministral-3-Reasoning |
| Cheapest dataset validation | SmolLM2 — validate data on a tiny model first, then scale |
Read the model contract in the UI before launching: it shows loader, VRAM, recommended instance, sequence length, batch size, and dependencies — the "will this fit?" answer is right there. The platform VRAM-checks before creating any endpoint (see 08-ship-deploy.md).
Real-world examples
🏦 Financial services — long context for contracts
A legal-ops team needs to reason over 100K-token contracts. They choose a long-context base — a choice they can't retrofit later — then fine-tune on their clause annotations.
🧑💼 SaaS — cheap enough to scale
A startup serving millions of chat completions picks a small MoE base so serving cost per request stays low at scale, then tunes tone via LoRA.
🏥 Healthcare — multimodal intake
A clinic wants models that can read a scan report image plus text. They pick a multimodal base — again, a locked-at-choice — and fine-tune the extraction style.
🛒 Retail — validate before you scale
A retailer with a brand-new dataset fine-tunes a small model on a sample first to confirm the data teaches what they expect, then scales to a larger base. Cheap insurance before spending on a big run.
Gotchas
- Context length and multimodality are locked at base selection. Style is fixable; context is not.
- The managed-model list is admin-controlled per purpose — if a model isn't in your selects, ask your admin (or see 16-model-governance.md).
- Architecture matters for serving cost: MLA/GQA keep long-context serving cheap; MoE gives more brain per FLOP.
consolidated.safetensors-style raw checkpoints and adapter-only tarballs are handled for you — you pick a model, not a file format.