# Models

> How to choose, pin, and discover models on sference: open-weight families, BYOM, and the catalog API.

Source: https://sference.com/docs/models

## Choosing a model

If you are **experimenting**, start with a small, fast open-weight checkpoint (for example **Qwen3.6 35B** or **DeepSeek V4 Flash** in the catalog). If you are **shipping to production**, optimize for **latency**, **cost per million tokens**, and **auditability**. sference is built around **pinned versions**, so the model behind an id never changes underneath you.

Every catalog model runs in all three [processing modes](https://sference.com/docs/guides/responses#processing-modes): **realtime** for user-facing traffic (where most usage is), **flex** for discounted latency-tolerant sync work, and **async** (24h window) for background pipelines: document extraction, synthetic data, embeddings backfills, and eval harnesses. Pick the smallest model that clears your quality bar, then scale horizontally.

> **Always pin**
>
> Model ids in requests should include the **exact catalog string** you intend to run. Silent upgrades from “latest” are hostile to compliance reviews; we treat pinned versions as part of the audit trail.

## Frontier vs efficient tiers

Think in three bands (names vary by catalog):

| Band | When to use |
| --- | --- |
| **Frontier-class open weights** | Hardest reasoning, codegen, or multimodal extraction where quality dominates unit cost. |
| **Mid-tier** | Production workloads where you still need headroom but not the largest context. |
| **Efficient / small** | High-volume async, classification, tagging, and preprocessing; optimize $/1M tokens first. |

## Live catalog and serverless rates

The full catalog with exact ids and current serverless rates (USD per 1M tokens, realtime priority). Every model is addressed by its exact id string, and `GET /v1/models` returns the live list your key can reach, including per-model pricing and capabilities. The same catalog, search, and deployment-request routes live in **API Reference** under **Model Catalog**.

| Model | Context | Input | Output | Cached input | Vision |
| --- | --- | --- | --- | --- | --- |
| `zai-org/GLM-5.2` | 1M | $1.20 | $4.20 | $0.26 | — |
| `zai-org/GLM-5.3` | 1M | $1.20 | $4.20 | $0.26 | — |
| `zai-org/GLM-5.3-Flash` | 1M | $0.20 | $0.60 | $0.07 | ✓ |
| `deepseek-ai/DeepSeek-V4-Flash` | 1M | $0.28 | $0.56 | $0.07 | — |
| `deepseek-ai/DeepSeek-V4-Flash-0731` | 1M | $0.28 | $0.56 | $0.07 | — |
| `bottlecapai/ThinkingCap-Qwen3.6-27B` | 256K | $0.40 | $2.60 | $0.05 | — |
| `Qwen/Qwen3.6-35B-A3B` | 256K | $0.20 | $1.25 | $0.05 | — |
| `Qwen/Qwen3-VL-30B-A3B-Instruct` | 256K | $0.40 | $2.00 | $0.10 | ✓ |
| `moonshotai/Kimi-K3` | 1M | $3.00 | $15.00 | $0.45 | — |

**GLM 5.2 retires 2026-10-01**, replaced by GLM 5.3 at identical rates. Coming soon: DeepSeek V4 Pro, Gemma 4, OpenAI GPT-OSS 120B, NVIDIA Nemotron 3 Super, Qwen3.8. Rates are confirmed per model at account provisioning; the site shows USD with an EUR toggle.

**Batch create** validates each row’s `model` against the catalog and rejects batches that mix models. Use one model id per `POST /v1/batches` job.

## Specialized paths

- [Batch inference](https://sference.com/docs/guides/batches): When you need JSONL-scale throughput and SLA windows instead of a single chat turn.
- [Responses & streams](https://sference.com/docs/guides/responses): Incremental output and event streams for long-running jobs.
- [Catalog & deployment](https://sference.com/docs/api-reference): Control plane routes for search, catalog, and requesting deployments.

## Bring your own model (BYOM)

If it runs on **vLLM** or **SGLang** in your stack, we can usually serve it on sference, subject to onboarding and capacity. Start from the **Model Catalog** and **request deployment** flows in **API Reference**, or reach out at [hello@sference.com](mailto:hello@sference.com) if you are still evaluating fit.

## Browse the full catalog

Use **API Reference → Control Plane → Model Catalog** for live schemas, or introspect from your integration tests. The marketing site also lists workload narratives under [use cases](https://sference.com/#use-cases).