Fireworks: Ember-1

Fireworks: Ember-1 is a text generation model from Fireworks, released by the provider on 24 September 2026. It is listed on api14 under the text category, and Fireworks describes it as a specialized reasoning model from Fireworks Research built on Kimi K3. The design goal stated by the provider is efficiency in reasoning: Ember-1 produces shorter reasoning traces, which Fireworks frames as making every token go further. For a developer, the practical implication is a model positioned around reasoning work rather than general chat alone, with reasoning exposed as a controllable part of the request.

Ember-1 accepts text and image input and returns text only, so it handles multimodal prompts where images are provided alongside text and the answer comes back as language. The context window is very large at 1,048,576 tokens, with a maximum output of 943,718 tokens. That combination of a million-token input budget and a large output ceiling makes the model relevant for workloads where long documents, large codebases or extended transcripts must be held in context while still producing substantial output in a single call.

The capabilities suggest a few concrete uses. Long-context reasoning over large inputs fits document analysis, multi-file code review, and summarisation or question answering across lengthy sources. Because images are accepted as input, text-and-image prompts such as reading diagrams, screenshots or scanned pages and reasoning over them are also supported. The reasoning controls and structured output support point towards agent-style workflows, tool calling, and extraction tasks where the model must return data in a defined shape. The ability to tune reasoning effort lets you trade depth of reasoning against response length depending on the task.

You call Ember-1 through api14's OpenAI-compatible API using the model id "fireworks/ember-1". Alongside the usual request fields, it supports frequency_penalty, presence_penalty, logit_bias, logprobs, top_logprobs, max_tokens, stop, temperature, top_p, top_k and repetition_penalty for sampling control. For reasoning, it supports reasoning, reasoning_effort and include_reasoning. For structured and tool-based work it supports response_format, structured_outputs, tools and tool_choice. If you only need plain text generation without these extended options, a simpler model may be enough, but Ember-1 is the fit when long context, image input or controllable reasoning are required.