Fireworks 已在其推理和训练平台上发布了 Kimi K3，这是一个拥有 3 万亿参数的开源模型。该模型拥有 100 万 token 的上下文窗口、原生视觉能力以及可与 Anthropic 和 OpenAI 的顶级闭源模型相媲美的推理能力。Kimi K3 的成本显著降低，并且可以选择在 Fireworks 上进行 LoRA 训练，允许用户针对特定工作负载微调模型。 AI
影响 此次发布提供了一个强大的开源替代方案，可与领先的专有模型相媲美，有望降低高级人工智能功能的成本并提高其可访问性。
排序理由 Kimi Moonshot 发布新的前沿开源模型 (Kimi K3)，并附有详细的技术规格和性能声明。
在 X — Fireworks (inference infra) 阅读 →
AI 生成摘要 · Google Gemini · 来自 8 个来源。 我们如何撰写摘要 →
完整方法见我们的编辑标准。
You can now fine-tune Kimi K3 on Fireworks. Conduct supervised fine-tuning, preference tuning, and reinforcement learning via Training API. Run across dedicated, and serverless training. The first open frontier model at 3 trillion parameters, ready for your product. Contact us
Start owning your own intelligence today with Kimi K3 on Fireworks. Serve it, fine-tune it, and put it into production. What are you waiting for → https://t.co/0yVyBZTAH1
Big props @Kimi_Moonshot on the Kimi Delta Attention (KDA) architecture that delivers a 2.5X better scaling efficiency. To unlock frontier performance, we put together: custom kernels for the KDA architecture, FP4 MoE kernels, and adapted decode kernels. Now it goes vroom.
Kimi K3 rivals Anthropic and OpenAI’s top models at a fraction of the cost. $3 / 1M input tokens. $15 / 1M output. $0.30 / 1M cached You can now route your hardest reasoning to an open model without paying that frontier premium.
Make Kimi K3 yours with LoRA training on Fireworks Training a model this massive used to be a big project. With Fireworks Training, you can take the best open model in the world and efficiently tune it to your unique workloads, without paying frontier prices or touching a GPU h…
Kimi K3 is live on Fireworks. Day 0, inference and training. US-hosted, and zero data retention. This is the first frontier open model in the 3 trillion parameter class. It sports 1M context, native vision, and reasoning that rivals the top closed models. Boom. https://t.co/ItU…
Kimi K3 on Fireworks AI brings frontier-level intelligence to an open-weight 2.8-trillion parameter model, delivering matching performance to closed alternatives like Opus 5 at up to 5x lower cost per task. Hosted with US-based serverless endpoints and Zero Data Retention (ZDR), …
Fireworks AI now supports private preview access for Serverless LoRA Training on Kimi K3, the world’s first open-weight 2.8-trillion parameter model. Through low-cost, pay-per-token serverless fine-tuning, developers can customize Kimi K3 directly from a laptop—eliminating comple…