Kimi K3 (Fireworks AI)

On July 27, 2026, Moonshot AI officially released the full weights of Kimi K3 as an open-weight model on Hugging Face, GitHub, and ModelScope, available for free download, local deployment, and secondary development. According to the 36Kr report, Kimi K3 has approximately 2.8 trillion parameters—the world's first open- The piece places this release in business-strategy context, noting Moonshot AI's ARR surpassed $300 million in mid-June 2026 with API revenue accounting for over 70% of that figure, that new consumer-end subscriptions were suspended on July 19, and that Moonshot closed its previous funding round at a $20 billion valuat

Harvey introduced Harvey Tenet, a legal-domain model built by post-training a Kimi K3 base with Fireworks through asynchronous reinforcement learning targeted at long-horizon legal agent work. The training corpus combined synthetic data, publicly available legal data, and human-expert data, and Harvey states that no cu The news connects the Kimi K3 model family to a specialized enterprise application in the legal sector, showing Fireworks' infrastructure being used for asynchronous RL post-training rather than only serving. This is a third-party derivative of Kimi K3 rather than a new Moonshot K3 release, and specific fine-tuning met

GitHub's official changelog confirms that Kimi K3 is generally available in GitHub Copilot, hosted by GitHub on Fireworks AI, with pricing of $3 per 1M input tokens, $15 per 1M output tokens, and $0.30 per 1M cached input tokens. The model is rolling out to Copilot Pro, Pro+, Max, Business, and Enterprise plans across The changelog notes an August 6, 2026 temporary pause of the rollout to mitigate a GitHub Actions incident, followed by resumed rollout the same day. For Copilot Business and Copilot Enterprise, Kimi K3 is off by default and plan administrators must enable the Kimi K3 policy in Copilot settings before members can selec

Microsoft announced that Kimi K3, Moonshot AI's 2.8-trillion-parameter open-weight model, is deployable on Microsoft Foundry through Fireworks AI. The model uses a Mixture-of-Experts architecture with 896 experts, activating 16 per token, and supports a one-million-token context window with native multimodal capabiliti In benchmark results reported by Neowin, Kimi K3 scored 81.2 on FrontierSWE, ahead of GPT-5.6 Sol and Claude Opus 4.8, and achieved leading results on BrowseComp, SWE Marathon, AutomationBench, SpreadsheetBench 2, and OmniDocBench. The article also notes that Kimi K3 is not under the unrestricted Modified MIT License u

The Fireworks AI landing page for the Kimi lineup explicitly introduces Kimi K3 as an open-weights, 2.8 trillion parameter model released on July 27th, with a 1,048,576-token context window and Vision capability. The page attributes Kimi K3 to Moonshot AI (an Alibaba-backed company founded by Zhilin Yang), preserving c The page also presents concrete performance claims: Kimi K3 reportedly achieves up to 201 tokens per second inference (3x faster than unnamed closed-source models and 4.5x faster than other providers) and reportedly outperformed "Fable" on 1,000 agentic tasks spanning security, crypto, and long terminal loops, at up to