Moonshot AI Just Released a 2.8 Trillion‑Parameter Open‑Source Model — And It Runs on AMD GPUs

On July 27, 2026, Moonshot AI released the model weights, technical report, and three infrastructure codebases (MoonEP, FlashKDA, and AgentEnv) for Kimi K3. Below is a breakdown, based on verifiable data and public documentation, of the actual implications across four dimensions.

1. Model Specifications and Benchmark Results

Kimi K3 has a total of 2.8 trillion parameters (2.8T) and activates approximately 104 billion parameters per inference (MoE architecture with 896 experts, selecting 16). The checkpoint size is 1.5609 TB (MXFP4 format).

Third‑party evaluations:

• WebDev Arena: Elo 1679, ranking 1st among open‑source models on that leaderboard, and 4th overall (including closed‑source models).

• Artificial Analysis Intelligence Index v4.1: composite score of 57.1, placing 4th globally, behind Claude Fable 5 (score undisclosed), GPT‑5.6 Sol (score undisclosed), and one other closed‑source model.

On the architectural side, the technical report discloses the following non‑scaling‑only changes:

• Attention mechanism: a 3:1 mix of KDA and Gated MLA.

• Expert stability: Stable LatentMoE, which keeps activation gradients stable across the 896 experts.

• Visual encoder: MoonViT‑V2, enabling native multimodal input.

2. Open‑Source Scope and License

In addition to the weights, three infrastructure components are released concurrently:

• MoonEP: a communication library for MoE expert parallelism, addressing cross‑node transfer bottlenecks.

• FlashKDA: an optimized kernel for KDA attention, delivering 1.72×–2.22× faster prefill on NVIDIA H20 (per the report).

• AgentEnv: a sandbox environment for large‑scale agent training.

The license is a modified MIT, allowing commercial modifications and closed‑source derivatives.

3. Hardware Deployment Costs: Measured and Estimated

AMD verified on day one that an 8‑GPU MI355X server (288GB HBM per GPU, 2.304 TB total VRAM) can load K3 and run with a 16K context window. This configuration did not report generation speed, time‑to‑first‑token, or concurrency limits.

Hardware cost estimates (based on public market quotes):

• 8×MI355X server: approximately $250,000–$350,000 USD (≈ RMB 1.7–2.4 million).

• Moonshot AI’s official technical documentation recommends, for production‑grade services, using 64 or more accelerators within the same high‑speed interconnect domain.

• 64 MI355X (8 servers of 8 cards): hardware investment roughly RMB 13.6–19 million; power demand near 100 kW, with annual electricity costs around RMB 0.8–1.0 million at industrial rates.

Alternative options (sufficient VRAM on paper, but unverified):

• 16×H200 (2.256 TB) requires two servers, costing about RMB 4.35–5.7 million.

• 16×B200 (2.88 TB) also needs two servers, roughly RMB 5.4–6.8 million.

• 24×H100 (1.92 TB) has insufficient VRAM margin; cloud rental for one year is estimated to exceed RMB 8.7 million.

AMD has not yet published performance validation at a 1M‑token context. The current 16K test only demonstrates that the model “can run” and does not constitute a production‑grade reference.

4. Business Path: Self‑Hosting vs. API

On the same day as the open‑source release, Together AI, Tencent Cloud, and Alibaba Cloud announced they would offer Kimi K3 API access. Under a pay‑per‑token model, for enterprises with monthly usage below 1 billion tokens, self‑hosting a 64‑card cluster—considering depreciation and electricity—is far more expensive than API spending (exact API pricing has not been disclosed, but cloud providers typically follow a pay‑as‑you‑go model).

Moonshot AI has not disclosed its API profit margin, but the combination of open‑sourcing weights and simultaneously launching cloud services aligns with the industry standard of “free base model, paid hosted service.”

5. Categorized User Groups

Given the current hardware and software validation status:

• Individual developers: cannot afford the RMB 1.7 million minimum hardware investment, nor do they have the operational capability for 64‑card interconnects.

• Small and medium‑sized enterprises: self‑hosting is uneconomical; API calls are the only feasible entry point.

• Large institutions or national‑level projects: can acquire clusters of 100+ cards and use K3 as a foundation for fine‑tuning or continued training.

As of July 28, 2026, no public case has demonstrated that 8×MI355X can run stably with contexts longer than 16K or under high concurrency.