Kimi K3 Triple Open-Source Release: 2.8T MoE + AgentENV Distributed Training + DSpark 423 tok/s

On July 27, 2026, Moonshot AI released everything for 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, 16 activated / 896 experts per token — the largest openly released MoE to date
• 1M-token native context window, native vision understanding (MoonViT3d vision tower)
• Architecture: 69 Kimi Delta Attention (KDA) linear-attention layers + 24 MLA layers interleaved, paired with an 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
• The activation function was replaced with the in-house SiTU (replacing SwiGLU); no existing MoE kernels worked, so teams like LMSYS and RadixArk had to write them from scratch

The architecture and benchmark results were announced on July 16; weights followed on July 27. SGLang, Miles, Modal, and Baseten all offered Day-0 support.

2. Benchmarks: Winning on Agents and Coding

K3 sits in the top tier on agentic, get-things-done benchmarks:

| Benchmark | K3 Score | Comparison | |------|---------|---------| | Frontend Code Arena | 1679, 1st | 48 points above Fable 5, 1st in 6/7 domains | | BrowseComp | 91.2 | Beats Sol 90.4 and Fable 5 88.0 | | AutomationBench | 53% / 30.8 | 1st | | SWE Marathon | 42.0 | Fable 5 only 35.0 | | Program Bench | 77.8 | 1st | | DeepSearchQA (F1) | 95.0 | 1st | | SpreadsheetBench | 34.8 | 1st | | FrontierSWE | 81.2 | 10 points above Sol | | Terminal Bench 2.1 | 88.3 | Only 0.5 behind Sol 88.8 |

Artificial Analysis composite intelligence index: 57, fourth place, behind Fable 5 (60) and two GPT-5.6 Sol configs (59). Moonshot 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 placed Fable 5 ahead of itself (76.9 vs 72.9) — an unusual posture among Chinese model launches.

3. AgentENV: Large-Scale Agent RL Training as Open-Source Infrastructure

Released alongside the weights, AgentENV (kvcache-ai/AgentEnv) is the hidden protagonist of this open-source drop. It is a distributed environment system for large-scale Agent training, powering K3's Agent RL training, 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% within 12 hours (DAPO math, no KL term)
• Trainer / inference engine co-located with mutually exclusive sleeping: base weights never move; KV cache + CUDA graphs are freed while the trainer works
• Adapter synchronization: only BF16 LoRA adapters (2,800 tensors) are transferred, saving ~48 GiB/GPU of transient IPC memory
• Verification stack: Schulman k3 KL estimator (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 GitHub repo — no prior open-source project could fully reproduce Agent RL training stacks at the K2.5/K2.7 scale.

4. Inference Performance: DSpark Pushes Single-Stream Decoding to 423 tok/s

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

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

Two engineering highlights worth a closer look:

1. Hybrid KDA memory management: KDA states are fixed-size and overwritten in place, contrasting with 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 the raw inputs Sᵢ = (vᵢ, kᵢ, gkᵢ, βᵢ) ≈ 1 KB are stored, then replayed by a single fold kernel after verification — passing bit-identical checks

5. Pricing and Barriers to Entry

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

6. Things Worth Remembering

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

• A KDA + MLA hybrid architecture runs at 3T scale in an open model for the first time, and Attention Residuals is a new design worth studying
• AgentENV turns large-scale Agent training stacks from closed-door engineering into open infrastructure — a game-changer for small and mid-sized teams pursuing Agent RL
• MXFP4 + LoRA synergy natively supported — quantized base weights stay frozen while BF16 LoRA adapters are applied on top of the quantized base GEMMs, significantly lowering the barrier to domain-specific fine-tuning of K3
• Moonshot itself wrote that without open-sourcing, K3 "could become the next Fable 5" — the shadow of the 19-day US export ban lingers, and open weights are a structural hedge

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

---

References

• SGLang and Miles Day-0 support for Kimi K3
• Kimi.ai X: Model weights release
• Kimi.ai X: AgentENV open source
• Kimi.ai 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 tops Claude and GPT on multiple benchmarks