Use Fireworks AI Models in Big-AGI.

Open-weights model served on Fireworks AI.

Open-weights model served on Fireworks AI.

Official release of DeepSeek V4 Flash, superseding the preview, with substantially enhanced agentic capabilities. Ships with a speculative decoding module atta…

Moonshot AI 2.8T-parameter flagship on Kimi Delta Attention, with native visual understanding and a 1M-token context for long-horizon coding and reasoning.

Fast serving path for Kimi K3: same model and quality, lower latency, higher per-token price.

Thinking Machines Lab first open-weights model: a 975B MoE (41B active) trained natively across text, image, and audio, with controllable thinking effort.

Fine-tuned adapter served on Fireworks AI.

Fine-tuned adapter served on Fireworks AI.

Fine-tuned adapter served on Fireworks AI.

Fine-tuned adapter served on Fireworks AI.

Fine-tuned adapter served on Fireworks AI.

Fine-tuned adapter served on Fireworks AI.

Fine-tuned adapter served on Fireworks AI.

Fine-tuned adapter served on Fireworks AI.

Fine-tuned adapter served on Fireworks AI.

Fine-tuned adapter served on Fireworks AI.

Fine-tuned adapter served on Fireworks AI.

Z.ai flagship with 1M-token context and multi-effort coding for long-horizon agentic tasks. New IndexShare architecture and improved MTP layer cut per-token co…

Fast serving path for GLM 5.2: same model and quality, lower latency, higher per-token price.

Coding-focused agentic model built on Kimi K2.6, with better end-to-end completion on long-horizon software engineering and ~30% fewer thinking tokens.

Fast serving path for Kimi K2.7 Code: same model and quality, lower latency, higher per-token price.

MiniMax 428B MoE (23B active) with Sparse Attention for efficient long context, tuned for long-horizon agentic coding and cowork.

Alibaba flagship closed model, available outside Alibaba infrastructure exclusively through Fireworks AI.

NVIDIA frontier-scale hybrid LatentMoE (550B params, 55B active) interleaving Mamba-2 and MoE layers, for multi-step agents and long-context reasoning.

Streamlined DeepSeek open MoE tuned for low-latency, high-throughput inference at 1M-token context, retaining most of Pro reasoning and coding quality.

DeepSeek flagship open MoE (1.6T params) for frontier reasoning, coding, and long-context work up to 1M tokens. Hybrid attention keeps long contexts efficient.

Moonshot AI native-multimodal agentic model tuned for long-horizon coding, autonomous execution, and swarm task orchestration.

Fast serving path for Kimi K2.6: same model and quality, lower latency, higher per-token price.

MiniMax MoE built for complex agent harnesses and elaborate productivity tasks, leveraging Agent Teams, Skills, and dynamic tool search.

Z.ai 754B-parameter MoE built for agentic engineering, with strong coding and sustained performance across long multi-round tasks.

Open-weights model served on Fireworks AI.

Open-weights model served on Fireworks AI.

Model served on Fireworks AI.

Model served on Fireworks AI.

OpenAI open-weight model for high-reasoning, agentic, general-purpose use that fits on a single H100.

OpenAI smaller open-weight model for lower-latency, local, and specialized use cases.

DeepSeek flagship open MoE (1.6T params) for frontier reasoning, coding, and long-context work up to 1M tokens. Hybrid attention keeps long contexts efficient.