The model
Inferam currently serves DeepSeek-V4.1-Flash-1017, an open-weights model, on Inferam’s own hardware. The served model can change; /v1/models always has the current ID.
| Model ID | DeepSeek-V4.1-Flash-1017 |
| Context window | 262,144 tokens (input plus output) |
| Default output limit | 32,768 tokens |
| Maximum output limit | 65,536 tokens |
| Tool calling | Yes, including parallel calls |
| Structured output | json_object and json_schema |
| Image input | Yes, as base64 data |
| Prompt caching | Automatic |
Context window
Section titled “Context window”A request’s prompt plus its output limit must fit in 262,144 tokens. A longer one gets a 400 with the model server’s message, for example “This server’s maximum context length is 262144 tokens, but the rendered prompt has 270112 tokens…”. Coding agents set up with the CLI compact their conversation before that. If you configured an agent by hand before 30 October 2026, when the window was 524,288, change it to 262,144 (see your agent’s page).
Reasoning
Section titled “Reasoning”The model thinks before answering by default, and its reasoning is returned separately from the answer. Each format has its own way to set the effort:
| Format | Parameter | Values |
|---|---|---|
| Chat Completions | reasoning_effort |
none, minimal, low, medium, high, xhigh, max |
| Responses | reasoning.effort |
the same |
| Messages | thinking, output_config.effort |
{"type": "disabled"} turns it off; budget_tokens caps it; effort low to max |
none (or "thinking": {"type": "disabled"}) turns thinking off entirely. minimal is treated as low. Reasoning tokens count toward the output limit, so on hard tasks leave room for them.
On tool-calling turns, send the model’s earlier reasoning back with the assistant message: reasoning_content in Chat Completions, reasoning items in Responses, thinking blocks in Messages. Coding agents do this already.
Images
Section titled “Images”The model accepts images as base64 data: URLs: an image_url part in Chat Completions, input_image in Responses, or an image block with a base64 source in Messages. Image URLs that point anywhere else are rejected with a 400, because Inferam never fetches URLs on your behalf. Audio and video aren’t supported.
Throughput
Section titled “Throughput”Generation runs at roughly 75 to 110 output tokens per second for a single stream, and prompt processing at about 2,000 tokens per second. The service runs a few generations at once. When all are busy, requests queue fairly for up to 30 seconds.
Known limitations
Section titled “Known limitations”- Quantised weights. The model runs as a 2.9-bit-per-weight quantisation to fit Inferam’s hardware. No published evaluation compares it with the full-precision model.
- No server-side tools. Web search, code execution and other provider-hosted tools aren’t available. Function tools you define work normally.
- No conversation state. Every request carries the whole conversation; Responses’
previous_response_idisn’t supported. - Token counting is an estimate.
/v1/messages/count_tokensapproximates the count; the usage in each response is exact. - Single region. The model runs in one location. Expect added network latency far from it, and see status.inferam.ai during incidents.