Model settings update: larger output budgets and corrected context windows
We corrected the context windows and maximum output budgets for the DeepSeek, GLM, Kimi, Qwen and MiMo lanes. The gateway now reads these limits from the model registry on every request, so large max_tokens requests are no longer clamped by a stale default.
We corrected the context windows and maximum output budgets for the main chat lanes. This was a configuration issue, not a model change: the gateway now reads the limits from the model registry on every request instead of falling back to a single default.
The practical effect is that agents and coding tools can ask for the output budget the model actually supports. A request for max_tokens: 131072 on DeepSeek V4.1 Flash 1017 now passes through unchanged. Previously it was clamped to 65,536 because the request path used a stale default.
Current limits
| Model | Context | Max output |
|---|---|---|
| DeepSeek V4.1 Flash 1017 | 1,048,576 | 393,216 |
| DeepSeek V4.1 Flash | 1,048,576 | 384,000 |
| DeepSeek V4 Flash 0731 | 1,048,576 | 131,072 |
| GLM 5.2 | 262,144 | 131,072 |
| GLM 5.3 | 327,680 | 131,072 |
| GLM 5.3 Flash | 327,680 | 131,072 |
| Kimi K3 | 1,113,088 | 1,048,576 |
| Qwen 3.8 27B | 1,001,536 | 32,768 |
| MiMo V2.6 Pro | 1,048,576 | 131,072 |
Output plus input cannot exceed the context window. The max output number is the ceiling for a single response, not a promise that every response will use it.
What changed
- The request normalizer now reads
context_windowandmax_output_tokensfrom the live model registry. - Requests above a model's output ceiling are clamped to that ceiling, not rejected.
- Requests within the ceiling are forwarded without modification.
- The live catalog at GET /v1/models is the source of truth. The model page at /models shows the same values.
Why it matters for agents
Long-running coding agents often send a large max_tokens even when the response will be short. A stale low ceiling makes the tool look broken or truncates a long file edit. If you are wiring an agent, read the limit from /v1/models and pass it through. Agent-assisted accounts can use the same limits after the first top-up. See /for-agents.
If you see a clamp_max_tokens event now, the requested value is above the published ceiling for that model. Check the model row before changing your client.
James Miano
CTO & ML Engineer at Lacesse Ventures. James builds agent-native inference infrastructure, pinned model routing and terminal-first API access.
Stop paying for overpriced round-trip latency
Why settle for slow routes and opaque pricing? Fikra pins DeepSeek checkpoints, publishes every rate, and lets agents create accounts, pay and work from a terminal. No minimum credit limits. Pay with card or mobile money.