Kimi releases K3 open model with 2.8T parameters and 1M context
TL;DR
Kimi launches K3, a multimodal open-weight model with 2.8 trillion parameters and 1M token context. It nears GPT-5.6 Sol and Claude Fable 5 in benchmarks, with full weights due by July 27.
What changed
Kimi released K3, a multimodal open-weight model with 2.8 trillion parameters and one million tokens of context. Developers and Vibe Builders can access it now while full weights arrive by July 27. The model runs pricier than its predecessor yet matches performance levels of Claude Fable 5 and GPT 5.6 Sol in company tests.
Specs
- •Parameters 2.8 trillion
- •Context window 1M tokens
Why it matters
Basic Users gain a stronger option for handling long documents and mixed media tasks than GLM 5.2 offered before. Developers see an open-weight path that reaches Claude Fable 5 territory without closed APIs. Vibe Builders can test generation quality against GPT 5.6 Sol benchmarks on creative multimodal projects.
What to watch for
Compare outputs directly with Claude Fable 5 on shared prompts. Developers should run their own evals on context-heavy tasks and monitor pricing shifts once weights drop.
Who this matters for
- Vibe Builders: Test K3 on creative multimodal projects to compare its generation quality against GPT 5.6 Sol.
Harsh’s take
Kimi releasing a 2.8 trillion parameter open-weight model is a massive technical milestone, but the era of dirt-cheap Chinese API subsidization is clearly drawing to a close. Builders must prepare for realistic infrastructure costs as performance parity with frontier models like GPT 5.6 Sol becomes the norm. Do not rely on venture-backed pricing models to sustain your margins. Focus on running local evals before the July 27 weight release to see if the 1M token context window actually holds up under heavy production loads.
by Harsh Desai
About Modal
View the full Modal page →All Modal updatesGo deeper
More AI news
- Daily RoundupQwen3.8-Flash-Next-Uncensored-GGUF trends, oMLX speeds agents, and Superagent for daily runs
A GGUF multimodal model climbed Hugging Face charts while two Product Hunt tools targeted faster local agents and simpler agent hosting on 30 August.
- Daily RoundupTiel-Coder-35B and GLM-5.3 top Hugging Face, MiniMax H3 on Fal, and DeepSeek V4 batch (new models for daily use)
Hugging Face saw five models trend including Tiel-Coder-35B-A3B-GGUF and GLM-5.3 while Fal added MiniMax H3 Max video, Cohere released Parse 5, and DeepSeek V4 Flash 731 reached OpenRouter for batch coding work.
- Daily RoundupVercel eve agents and CLI updates, Hy4 on AI Gateway, and new model drops for agents
Vercel pushes agent deployment and CLI tools while new models from Tencent, unsloth, and others land on gateways and hubs for immediate testing and integration.