Kimi K3 Triple Open-Source Release: 2.8T MoE Model, AgentENV Distributed Training, and 423 tok/s DSpark Inference

On July 27, 2026, Moonshot AI released everything behind Kimi K3 at once: model weights, high-performance attention kernels, MoE communication libraries, and the AgentENV distributed training environment — all handed to developers on the same day. This is the first time the open-source world has seen a 3-trillion-parameter-class MoE.

1. What scale of model is Kimi K3?

• 2.8 trillion total parameters, with 16B/896B active parameters — the largest openly released MoE
• 1M-token native context window, native vision understanding (MoonViT3d vision tower)
• Architecture: 69 Kimi Delta Attention (KDA) linear-attention layers + 24 MLA layers interleaved, with 896-expert LatentMoE (routing in a 3584-dim latent space)
• Key engineering innovation: Attention Residuals — every attention output is banked and carried through the entire network stack, bypassing standard layer-pipeline assumptions
• Activation replaced with the in-house SiTU (instead of SwiGLU); no off-the-shelf MoE kernels existed, so teams like LMSYS and RadixArk wrote them from scratch

The architecture and benchmarks were announced on July 16; weights followed on July 27, with SGLang, Miles, Modal, and Baseten all offering Day-0 support.

2. Benchmarks: leading in agents and coding

K3 sits in the top tier on agentic, "can it get work done" benchmarks:

| Benchmark | K3 score | Comparison | |------|---------|---------| | Frontend Code Arena | 1679, first | 48 points above Fable 5, first in 6/7 domains | | BrowseComp | 91.2 | Ahead of Sol 90.4 and Fable 5 88.0 | | AutomationBench | 53% / 30.8 | First | | SWE Marathon | 42.0 | Fable 5 only 35.0 | | Program Bench | 77.8 | First | | DeepSearchQA (F1) | 95.0 | First | | SpreadsheetBench | 34.8 | First | | FrontierSWE | 81.2 | 10 points above Sol | | Terminal Bench 2.1 | 88.3 | Just 0.5 behind Sol's 88.8 |

Artificial Analysis' composite intelligence index puts K3 at 57, fourth overall — behind Fable 5 (60) and two GPT-5.6 Sol configurations (59). Moonshot itself was candid: "K3 does not beat Fable 5 on general intelligence — but it competes." Notably, on Moonshot's own Kimi Code Bench 2.0, K3 ranked Fable 5 *above itself* (76.9 vs 72.9) — an uncommon stance among model launches.

3. AgentENV: large-scale agent RL training as open infrastructure

Released alongside the weights, AgentENV (kvcache-ai/AgentEnv) is the hidden protagonist of this release: a distributed environment system for large-scale agent training, supporting Kimi K3's agentic RL, with:

• Fast snapshot / restore / branching, suited to massively parallel agent workflows
• Paired with the Miles RL framework, it pushed AIME-2024 (greedy) from 43.3% to 76.7% in 12 hours (DAPO math, no KL term)
• Trainer/deployment engine co-located with mutually exclusive sleep: base weights never move; KV cache + CUDA graphs are freed while the trainer works
• Adapter sync: only BF16 LoRA adapters are transferred (2,800 tensors), saving ~48 GiB/GPU of transient IPC memory
• Verification stack: Schulman k3 KL estimation (quantized lower bound ~2e-3) + canary sync probes + tensor-level dumps + SHA256 manifests

AgentENV turns "writing training environments for agents" from closed-door engineering into a public GitHub repository — no open-source project had previously enabled full reproduction of agent RL training stacks at the scale of K2.5 or K2.7.

4. Inference: DSpark pushes single-replica decode to 423 tok/s

The SGLang backend written by the LMSYS and RadixArk teams pushes K3's inference optimization to the limit:

• Batch-1 decode ~113 tok/s (no speculative decoding; 2×4 GB300 hardware)
• ~423 tok/s with DSpark speculative decoding
• Serving-frontier PP8→TP8 endpoint fp4 path at 2,808 tok/s/GPU
• Under agent traffic, DCP8 vs TP8 raises logical KV capacity by ~7.9× (1.5M → 12.2M tokens), with 541 tok/s at 48 concurrent sessions
• Draft window cut from 512KB to 16KB — roughly 32× memory reduction (ReplaySSM technique)

Two engineering highlights worth a closer look:

1. Hybrid KDA memory management: KDA state is fixed-size and overwritten in place, in contrast to append-only KV; a unified memory pool lets KDA state grow from one end and MLA KV blocks from the other (KDA blocks ~54 MB vs MLA KV blocks ~27 KB — three orders of magnitude apart) 2. KDA state restoration for draft steps: instead of snapshotting K×V state per step (64 KB/request/layer/head), only raw inputs Sᵢ = (vᵢ, kᵢ, gkᵢ, βᵢ) ≈ 1 KB are stored, followed by a single fold-kernel replay after verification — bit-identical verification passes

5. Pricing and requirements

• Hosted API: $3 / M input tokens, $15 / M output tokens — departing from the low-price playbook of Chinese open-source models and entering frontier-closed-model territory
• Minimum self-hosting bar: 1.4 TB aggregate GPU memory; realistically 8 nodes × 8× 80GB GPUs (5.12 TB, leaving headroom for KV cache and activations)
• MXFP4 quantization cuts memory bandwidth by ~4× vs FP16, making the cost structure viable for mid-sized organizations
• The 1M context window is exclusive to the Allegretto tier (from ¥199/month)

6. Things worth remembering

K3 is not "just another open-source MoE":

• The KDA + MLA hybrid architecture runs at 3T-class open-source scale for the first time, and Attention Residuals is a new, borrowable design
• AgentENV turns large-scale agent training stacks from closed-door engineering into open infrastructure — a leapfrog enabler for small and mid-size teams pursuing agent RL
• Native MXFP4 + LoRA synergy — quantized base weights stay frozen while BF16 LoRA adapters stack on top of quantized base GEMMs — significantly lowers the barrier to domain fine-tuning of K3
• Moonshot itself wrote that "without open-sourcing, we could become the next Fable 5" — with the shadow of a 19-day-old U.S. export ban still present, open weights are a structural hedge

Musk left a one-word "Impressive" under Artificial Analysis' tweet. Vercel CEO Guillermo Rauch ran nextjs.org's combined web-engineering eval and concluded "an open model leading all proprietary models for the first time." Sriram Krishnan called it "a major moment with multiple implications for the entire industry," and Emad Mostaque was more blunt: "U.S. labs will eventually end up distilling Chinese models."

---

References

• SGLang and Miles provide Day-0 support for Kimi K3
• Kimi.ai on X: model weights release
• Kimi.ai on X: AgentENV open-sourced
• Kimi.ai on X: Modal Day-0 DFlash speculative decoding
• Kimi K3 model overview: MXFP4 quantization and open-source significance
• 36Kr: Musk "Impressed" by K3
• Tencent News: 2.8T-parameter K3 beats Claude and GPT on multiple metrics