Enable, test, and select models

After 0G Compute is configured and the account has funds, enable at least one primary model for Chat or Coding. Only models with their switch on appear in the primary Chat and Coding pickers; the Background Model is selected independently under Settings → General.

Where models come from

Ghast does not ship a permanent model list or read separate catalogs from every model company. Settings → 0G Compute loads currently available models and serving-provider data from 0G Router.

UI term	Meaning
0G Compute	The current user-configurable model provider and routing entry point in Ghast.
OpenAI, Anthropic, DeepSeek, Qwen, and others	Model brands or families in the 0G catalog, not separate Ghast providers to configure.
1P, 2P, 3P	The current number of providers reported by 0G for that model.
TeeML, TeeTLS, TDX, and similar labels	Verifiability or trusted-execution metadata returned by the 0G catalog; use the current hover text for the exact meaning.

One model brand can be served by one or more providers behind 0G Router. Ghast does not currently expose a control for pinning one serving provider; model requests use the Router path prepared for that catalog entry. Catalog entries, provider count, price, and availability can change, so the live page takes precedence over screenshots and examples.

Which models are available

The catalog is grouped by family. This is the catalog snapshot observed in the current development client; it describes that live capture, not a fixed quota:

Family	Current count	Family	Current count
DeepSeek	3	Qwen	27
GLM	8	Kimi	5
MiniMax	4	Tencent	2
0GM	2	Anthropic	4
OpenAI	5	Other	2

The total was All 62. Counts can change after a refresh.

The catalog can contain 0GM, DeepSeek, GLM, GPT, Hunyuan, Kimi, MiniMax, Qwen, and other models, including vision or video-generation entries. These current examples explain the structure and are not a promise of permanent availability:

• 0GM-1.0-35B-A3B
• DeepSeek-V4-Flash, DeepSeek-V4-Flash-Vision (experimental), DeepSeek-V4-Pro
• GLM-5, GLM-5.1, GLM-5.2, GLM-5.3-Flash
• GPT-5.6 Luna
• Hunyuan-3
• Kimi-K2.7-Code, Kimi-K3
• MiniMax-H3
• Qwen3-VL-30B-A3B-Instruct, Qwen3.6-Plus, and others

For the complete list at this moment, open Settings → 0G Compute and select Refresh. Do not use the examples above to conclude that an account must have a specific model.

A real section of the model catalog. Enabled models sort first; counts, prices, and provider counts change with the Router catalog.

Read a model row

Item	How to interpret it
Model name	The name shown in Ghast model pickers. experimental means the catalog marks the entry experimental.
In / Out	Text models normally show 0G cost per million tokens; cached-input price appears separately when available.
Video	Video models show per-second cost and are not ordinary text-chat models.
262K, 1M, and similar	Context length returned by the catalog. This is a limit, not a target to fill.
Capability icons	Can indicate image input, image output, Tool calling, reasoning, or another detected capability. Hover for exact text.
Verifiability / execution labels	For example TeeML, TeeTLS, or TDX. Treat these as current 0G catalog metadata and do not generalize beyond the visible explanation.
nP	Current provider count. More providers does not mean one request uses every provider.
Switch	Controls whether the model appears in the primary Chat and Coding pickers.
Lightning button	Runs a real short inference test for this model and reports the result.

1. Refresh the catalog

Open Settings → 0G Compute and select Refresh at the upper right of Models. A refresh can change models, price, context, capabilities, and provider count.

If it fails, keep the visible error and check the Management Key, inference API key, account session, and Settings → Network. Do not rotate keys after a single catalog timeout.

2. Enable a model

1. Use the family filters to narrow the list.
2. Compare capabilities, context, input/output price, and provider count.
3. Turn on the switch at the right of the model row.
4. The model moves into the enabled-first section and appears in the primary Chat and Coding pickers.

The model switch saves immediately; it does not wait for Settings Save. Allow the state to settle, then reopen the conversation's model picker to confirm the model is available. If saving fails, the switch can revert without a separate error notification. Check the connection and try again after resolving the issue, rather than repeatedly toggling it.

Turning a switch off removes that model from the primary Chat and Coding pickers but does not delete conversations or alter previous responses. If a conversation had selected a model that is now disabled, choose an available model before the next send. This switch does not control the Background Model list.

3. Test a model

Select the lightning button at the far right of the row. It sends a real short request to that model rather than only probing a network port, so it can consume a small amount of inference usage.

• A loading state appears during the test.
• Success shows a success icon and latency, such as 1420ms.
• Failure shows No response; the accessible description or page error contains the failure reason.
• Video-generation entries do not currently show the same inline text-model test button.

A successful short test proves only that the model returned a short result at that moment. It does not validate long context, images, Tools, Skills, MCP, Coding commands, or external services.

4. Select the primary model in Chat or Coding

1. Return to a Chat or Coding conversation.
2. Open the model and reasoning menu at the lower right of the composer.
3. Open Model and choose from the enabled models grouped by family.
4. If the model exposes reasoning levels, open Reasoning and choose the level for the next run.

The selection applies to later generations and does not rewrite completed messages. Check Chat and Coding conversations separately. Switching models does not change Tool, Skill, MCP, or command authorization.

5. Set the Background Model

Open Settings → General → Background Model. Ghast uses it for activity summaries, query rewriting, memory organization, and similar background LLM work; it is not the primary Chat or Coding model.

The Background Model list reads the available 0G catalog and does not require the model-row switch to be on. A fast, lower-cost text model is usually appropriate. Select Test after changing it. A successful Background Model test does not validate the model selected in the current conversation.

Choose a first model

Choose by task, not only by brand:

Task	Check first
General questions and summaries	Text capability, price, latency, enough context
Coding and Tool workflows	Tool-calling capability, coding quality, context, connection test
Image understanding	Image-input or Vision marker
Long documents or large projects	Context limit, input price, compaction strategy
Video generation	Video model and per-second price; do not use it as an ordinary chat model

For a first session, enable only a small Chat and Coding set, such as a daily primary model and a stronger complex-task model. Then choose a lower-cost Background Model separately under Settings → General. Adjust when the live catalog changes rather than enabling everything.

When a model is unavailable

Symptom	Check
No model in the composer	Turn on at least one switch in 0G Compute, then reopen the picker.
Empty catalog	Refresh; check the Management Key, inference key, and network.
Test says No response	Check account funds, current provider count, network, and the visible error; retry once later.
Short test works but the real task fails	Check context, input type, Tool/Skill permissions, and task size.
Price or model differs from a screenshot	Use the current 0G Compute catalog.
The old model still seems active	Confirm the new name in the composer before sending; old messages do not change.

Done when At least one 0G Compute model is enabled and passes its test, Chat or Coding can select it, and Account funds refresh successfully.

Next: Quickstart. For deeper model-selection behavior, see Models and 0G Compute.