Kimi K3

Moonshot AI's 2.8-trillion-parameter, 104-billion-active, native multimodal agent model with a one-million-token context. Keep it as a serious task-specific model route; promote it only when Kevin's own harness receipts beat the incumbent on quality, cost, latency, refusals, privacy, and reliability. Source: Kimi K3 technical report and repository, captured 2026-08-10

Current snapshot

Source: Kimi K3 README, technical report, GitHub, and Hugging Face API, captured 2026-08-10

What is actually novel

K3 scales information flow in three directions: KDA mixes across long token sequences, AttnRes retrieves selectively across earlier layers, and Stable LatentMoE expands channel capacity while activating only a sparse subset of experts. Moonshot reports an approximately 2.5x overall scaling-efficiency gain over Kimi K2 from the combined architecture and training recipe. That is the report's own system-level comparison; the launch post's “up to 6.3x faster decoding” and “~25% higher training efficiency at <2% additional cost” remain component-level vendor claims until reproduced in an applicable serving or training setup. Source: Kimi K3 technical report §§2–3; X/@Kimi_Moonshot, 2026-07-16

The open checkpoint matters, but “open-weight” is the precise term. K3 uses a custom license rather than Apache, MIT, or another standard OSI license.

API contract

K3 always thinks and returns reasoning_content. The top-level reasoning_effort accepts low, high, or max, with max the default. For multi-turn conversations and tool calls, the client must pass the complete assistant message back unchanged—including reasoning_content and tool_calls—rather than retaining only content. Failing this is a harness compatibility bug, not a model-quality result. Source: Kimi K3 README §6

The official API supports structured JSON schemas, tool choice, dynamic tool loading, and image/video file input. The platform listed cache-hit, input, and output prices of $0.30, $3.00, and $15.00 per million tokens at review time. Recheck pricing and provider behavior before a run. Uploaded files and hosted prompts leave the local machine, even when extraction or storage is temporarily free. Source: Kimi API K3 quickstart, pricing, file, and video guides, reviewed 2026-08-10

Evaluation evidence

The two saved images are useful launch artifacts, not a standing leaderboard. They show K3 across coding, general-agent, and visual-agent suites, but the technical report mixes harnesses and sources: Kimi Code for K3, Claude Code or Terminus for some Claude/GLM runs, Codex for GPT runs, and cited third-party numbers elsewhere. The report itself says K3 trails the strongest proprietary systems overall. Source: saved launch media; Kimi K3 README evaluation notes

The durable evaluation rule is:

1. freeze the exact K3 endpoint or weight revision, harness, effort, context strategy, tools, and task fixture;
2. compare against the incumbent using the same final-state verifier and independent evidence;
3. record cost, latency, token use, refusals, tool-call validity, and failure taxonomy;
4. promote a route only when the task-specific receipt wins and can be replayed.

Security and Review Skills supplies one time-bounded, private route signal: a July benchmark favored K3 for continuous security-review price/quality. Its public benchmark snapshot later shifted, so Deepsec still requires a live 50-file calibration before changing the default. Source: Security and Review Skills, reviewed 2026-08-10

Harness signal

K3's most important contribution to this wiki is methodological. Its white-box RL environment represents an agent harness as configurable tool interfaces, system prompts, context management, skills, memories, subagents, and protocols. Training varies these modules across Kimi Code, Claude Code, Codex, OpenClaw, Hermes, and novel configurations to reduce single-harness overfit. This strongly supports Orchestration Variables and Harness Ablation Design: compare the full model–harness condition, then ablate one declared variable at a time. Source: Kimi K3 technical report §4.2.1

The report's knowledge-graph task synthesis is also useful, with a boundary. It expands coarse concepts into a deduplicated hierarchical DAG, samples a node and ancestor context, retrieves public material, and synthesizes a verifiable task. Kevin's Brain Source Fabric can borrow the frontier-expansion pattern, but synthesized tasks and graph edges remain derived candidates until source review; they never become evidence merely because an agent generated them. Source: Kimi K3 technical report §4.2.2

License and deployment gates

The Kimi K3 License permits use, modification, distribution, fine-tuning, and sale, subject to notice and legal-compliance terms. Two commercial thresholds matter:

• a Model-as-a-Service business above $20M aggregate affiliate revenue over any consecutive 12 months needs a separate Moonshot agreement before commercial use;
• a commercial product above 100M monthly active users or $20M monthly revenue must prominently display “Kimi K3.”

The thresholds do not apply to internal use or access through official/certified providers. This is an operational summary, not legal advice; re-read the frozen license before shipping. Source: Kimi K3 License at frozen HEAD 3cb39dfd

Self-hosting a 2.8T MoE is an infrastructure project, not a casual local-model install. Confirm the exact quantization, engine recipe, accelerator/RAM budget, concurrency, context target, telemetry, and license before downloading weights. No runtime, weights, credentials, or global skill were installed during this review.

FlashKDA is the current low-level KDA-kernel research route, not a general K3 speed switch. The saved launch reports 1.72–2.22x prefill speedup over a named flash-linear-attention baseline on H20 shapes. Current source is MIT at MoonshotAI/FlashKDA@1ce47ea3, with six detected test files. Reproduce supported GPU/driver/CUDA/CUTLASS shapes, numerical tolerances, forward/backward results, model-quality parity, build and fallback dispatch, warmup, memory, and end-to-end latency before adoption; do not project a kernel microbenchmark onto total model throughput. Source: X 2081762799202746420; commit-pinned repository receipt, 2026-08-12

Routing

Use K3 as a candidate, not the automatic default, when at least one of these is central:

• million-token context with explicit context-management evaluation;
• native image/video reasoning inside an agent loop;
• long-horizon coding, research, or knowledge work;
• open-weight deployment or architecture research;
• a controlled alternative to an incumbent model in The Eval Loop (Slop Is an Output Problem).

Start with a bounded hosted-API evaluation when the data-egress and spend rules allow it. Move to self-hosting only when open-weight control creates enough value to justify the hardware and operational burden. Use the existing harness-native model for ordinary coding until K3 wins the relevant local receipt.

Transparency limits

The report describes pretraining categories—web text, code, mathematics, knowledge, captions, interleaved image-text, OCR, perception, video, and visual coding—and filtering, deduplication, synthesis, and rephrasing. It does not provide a full dataset inventory, proportions, or source-by-source rights and provenance. Vendor case studies and benchmark tables should therefore remain attributed claims, while deployment decisions use independent and local proof.

---

Timeline

• 2026-08-12 | Added FlashKDA as a bounded kernel-research route at 1ce47ea3, MIT, with six detected tests. Its H20 microbenchmark schedules shape/numerics/build/fallback/end-to-end reproduction; it does not change the K3 model route. Source: X 2081762799202746420; repository evidence

• 2026-08-10 | Deep-reviewed the saved launch post, both benchmark images, current official repository, 47-page report, Hugging Face checkpoint, API contract, pricing, and custom license. Added a guarded K3 model route and promoted the white-box multi-harness and graph-guided frontier-expansion methods into their existing owners. No weights or runtime were installed. Source: X/@Kimi_Moonshot; MoonshotAI/Kimi-K3; Kimi API

• 2026-07-27 | Moonshot published the full K3 weights and repository under the custom Kimi K3 License. Source: GitHub and Hugging Face

• 2026-07-16 | Moonshot announced K3 through Kimi.com, Kimi Work, Kimi Code, and the Kimi API, with open weights promised by July 27. Source: X/@Kimi_Moonshot