LM Studio Bionic Adds Kimi K3, the Biggest Open-Weight Model Ever Built

LM Studio Bionic just added Kimi K3 to its cloud model lineup, giving users access to the largest open-weight model ever released without having to wrestle with self-hosting infrastructure. The model runs on US-based servers with Zero Data Retention (ZDR) enabled by default, meaning your prompts and outputs are never stored after a request completes. For anyone already using Bionic for coding or document work, it's a one-click model switch.

The biggest open-weight model ever shipped

Kimi K3 is a 2.8-trillion-parameter open-weight, multimodal reasoning model with a 1-million-token context window and always-on thinking mode. To put the scale in perspective: two weeks ago, the largest open-weight model in existence had one trillion parameters. Then Moonshot AI shipped Kimi K3, the first open model to reach 2.8 trillion parameters, nearly triple the size of its predecessor Kimi K2.6.

Moonshot's self-reported benchmarks have K3 mostly beating Claude Opus 4.8 max and GPT-5.5 high, while losing out to Claude Fable 5 and GPT-5.6 Sol. That puts it squarely in the frontier tier, even if it doesn't quite top the leaderboard. The key word here is open: this is the first time a model at this performance level has been available with public weights.

Under the hood: three architectural bets

The raw parameter count is the headline, but the engineering story is more interesting. Moonshot made three core architectural choices that make a 2.8T model actually serviceable:

• Kimi Delta Attention (KDA): KDA is a linear attention mechanism that replaces traditional quadratic attention. Standard attention scales quadratically with sequence length, which becomes prohibitively expensive at 1M tokens. KDA delivers 6.3x faster decoding at 1M context.
• Attention Residuals (AttnRes): A drop-in replacement for standard residual connections. Rather than uniformly accumulating layer outputs, AttnRes allows each layer to selectively retrieve representations from arbitrary earlier layers. This adds 25% training efficiency at just 2% compute cost.