Quick verdict: test Kimi K3 if you need Kimi’s newest long-horizon, multimodal, 1M-context model. Moonshot has now released the weights and serving materials under the Kimi K3 License. Call this open weights, not generically open source, and review the license and substantial deployment requirements before self-hosting.
K3 is a serious frontier-adjacent Kimi upgrade: Moonshot describes a 2.8T-parameter model with native vision, Kimi Delta Attention, Attention Residuals, 1M context, and strong launch-suite results. The current AIHackers position is narrower: K3 is worth a controlled eval, but the launch evidence is mixed by provenance. Moonshot’s benchmark table is vendor evidence, Artificial Analysis is an early independent signal, and AIHackers has not yet run a site-owned repository test.
Quick Facts
| Spec | Kimi K3 |
|---|---|
| Provider | Moonshot AI / Kimi |
| Model ID | kimi-k3 |
| Parameters | 2.8T total parameters, vendor-stated |
| MoE routing | 16 of 896 experts active, vendor-stated |
| Architecture notes | Kimi Delta Attention, Attention Residuals, Stable LatentMoE |
| Context window | 1M tokens |
| Input / output | Text and image input; text output in current independent listing |
| API reasoning | Always-on thinking; pay-as-you-go API currently supports reasoning_effort: max only |
| Kimi Code reasoning | Membership model ID k3 supports low, high, and max; access and 256K/1M context depend on plan tier |
| API pricing | $0.30 cache-hit input / $3.00 cache-miss input / $15.00 output per 1M tokens |
| Quantization notes | Moonshot says quantization-aware training uses MXFP4 weights and MXFP8 activations from SFT onward |
| Deployment note | Moonshot recommends supernode deployments with 64 or more accelerators and says vLLM KDA support is planned alongside the model |
| Weights | Released in Moonshot’s official repository under the Kimi K3 License |
Kimi’s current platform model list also says kimi-k2.5 and the moonshot-v1 series are closed to new registered users after K3’s release, with full platform sunset on August 31, 2026. Treat that as a migration clock for older Kimi integrations, not as proof that K3 is the cheapest replacement.
Benchmark Evidence
| Source | Kimi K3 signal | How to read it |
|---|---|---|
| Moonshot launch blog | K3 max reasoning is shown near GPT-5.6 Sol, Opus 4.8, GLM-5.2, and Fable 5 across coding, agentic, reasoning, and vision tables | Vendor-reported launch suite; benchmark harnesses and fallback behavior vary |
| Artificial Analysis | Intelligence Index 57, 1M context, 62 output tokens/s, $3 input / $15 output | Early independent shortlist signal; not an AIHackers repo eval or license analysis |
| AIHackers | Not verified | No accepted-task field test has been run yet |
benchmark artifact
Kimi K3 Evidence Snapshot
| Model | Provider | Status | Context | Input price | Output price | Coding signal | Tool-use signal | Benchmark evidence | Speed | Verdict | Sources | Checked |
|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Kimi K3 | Moonshot AI | active Active Kimi API/product flagship; official weights and serving materials released under the Kimi K3 License. | 1M | $3.00 / 1M | $15.00 / 1M | Moonshot reports strong max-reasoning launch-suite coding and agent results; AIHackers repo eval is not verified. | Kimi API and Kimi Code support K3; Kimi docs warn to preserve full assistant history and avoid mid-session model switching. |
| Artificial Analysis reports 62 output tokens/s and flags high verbosity; measure total output cost per accepted task. | Test as Kimi's newest 1M-context frontier-adjacent lane; keep K2.7 Code for cheaper routine Kimi coding until K3 passes local CAR tests. | Kimi K3 launch blog [archive], Kimi K3 quickstart [archive], Kimi K3 API pricing [archive], Kimi K3 weights and license [archive], Kimi current model list [archive], Kimi Code model configuration, Artificial Analysis: Kimi K3 [archive] | 2026-08-01 |
| GLM-5.2 | Z.AI | active Current Z.AI flagship coding model and supported-tool value lane. | 1M | $1.40 / 1M | $4.40 / 1M | Z.AI reports 62.1 on SWE-Bench Pro and 81.0 on Terminal-Bench 2.1. | Supported-tool coding lane; BFCL score not imported. |
| Artificial Analysis flags higher output-token use; measure total cost per successful task. | July value pick to test for supported coding-tool workflows; keep Opus/GPT for final arbitration until local evals pass. | Z.AI GLM-5.2 overview [archive], Z.AI pricing [archive], Artificial Analysis: GLM-5.2 article [archive], Artificial Analysis Intelligence Index v4.1, SWE-bench, Berkeley Function Calling Leaderboard | 2026-06-28 |
| Kimi K2.7 Code | Moonshot AI | active Cheaper routine Kimi coding API lane; HighSpeed is the same model at higher token prices. | 256K | $0.95 / 1M | $4.00 / 1M | Kimi K2.7 Code remains the lower-cost 256K coding lane after K3; independent normalized benchmarks are not imported. | OpenAI-compatible API; thinking mode required in the documented K2.7 Code quickstart. |
| HighSpeed model ID exists at a higher token price; latency not independently measured here. | Cheaper routine Kimi coding API lane when Kimi routing fits and 256K context is enough. | Kimi K2.7 Code quickstart [archive], Kimi K2.7 Code pricing [archive], Kimi Code K2.7 release notes [archive], SWE-bench, Berkeley Function Calling Leaderboard | 2026-06-28 |
| Claude Opus 4.8 | Anthropic | historical Still available, but superseded by Opus 5 for current premium comparisons. | 1M | $5.00 / 1M | $25.00 / 1M | Historical premium Claude baseline; use Opus 5 for new task-level comparisons. | Still available for pinned integrations; new Claude premium routing should test Opus 5. |
| Artificial Analysis measured 57.3 output tokens/s; provider and workload latency vary. | Historical premium baseline. Use Claude Opus 5 for current Claude premium routing. | Claude models overview [archive], Claude API pricing [archive], Artificial Analysis: Claude Opus 4.8 [archive], Artificial Analysis Intelligence Index v4.1, LMArena leaderboard dataset, Berkeley Function Calling Leaderboard | 2026-07-25 |
| GPT-5.6 Sol | OpenAI | active Generally available through ChatGPT paid plans, Codex paid plans, and the OpenAI API; plan and effort options vary. | 1.05M | $5.00 / 1M | $30.00 / 1M | Artificial Analysis reports 80 on its Coding Agent Index at max effort; OpenAI reports 64.6% on SWE-bench Pro. | Generally available in API and paid Codex plans; max and ultra modes are vendor-documented. |
| OpenAI announced a selected-customer Cerebras preview for July; production latency is not verified. | Generally available flagship; test on real tasks and apply stronger controls for agentic or cyber work. | OpenAI GPT-5.6 general availability [archive], OpenAI API pricing [archive], Artificial Analysis GPT-5.6 evaluation [archive], Agent Arena leaderboard, OpenAI GPT-5.6 availability [archive], OpenAI GPT-5.6 system card | 2026-08-01 |
Kimi K3 has vendor launch evidence, an early Artificial Analysis row, and released weights under the Kimi K3 License. AIHackers has not yet run a controlled repository evaluation.
Do not convert Moonshot’s launch table—or one independent aggregate—into a blanket task-level winner claim. In the current Artificial Analysis v4.1 snapshot, K3 is 57 at 62 tokens/s and Opus 4.8 is 56 at 57.3 tokens/s. That is a shortlist signal under one benchmark mix, not proof that K3 will produce better accepted patches, reviews, or production outcomes.
Kimi K3 vs Current Alternatives
| Alternative | Source-labeled comparison |
|---|---|
| GPT-5.6 Sol | Moonshot compares K3 against Sol in its launch suite. GPT-5.6 is now generally available, but the cross-vendor launch comparison remains vendor evidence. |
| Claude Opus 4.8 | Opus remains the practical premium Claude baseline. K3 is cheaper on listed output price but not proven cheaper per accepted patch. |
| GLM-5.2 | GLM-5.2 keeps the low-cost 1M-context value lane at $1.40/$4.40. K3 costs more and should be tested as a stronger escalation, not a default replacement. |
| Kimi K2.7 Code | K2.7 Code remains the cheaper routine Kimi coding lane at 256K context. Route latest-Kimi and 1M/multimodal frontier intent to K3; route budget Kimi coding intent to K2.7 Code. |
The practical switch is not just an ID swap. K3 changes context, reasoning behavior, price, cache behavior, and migration risk.
Production Caveats
- Preserve the full assistant history. In multi-turn and tool-calling API flows, pass the complete prior assistant message back unchanged, including thinking/reasoning fields—not only the final
content. Moonshot warns that dropping this history can make K3 unstable. - Do not switch to K3 mid-session. K3 was trained with preserved thinking history. Start a new session instead of moving an active K2.x or other-model conversation to K3.
- Constrain proactivity explicitly. For bounded production agents, state approval boundaries, permitted actions, and stop conditions in the system prompt or
AGENTS.md; Moonshot warns that K3 may otherwise improvise around ambiguity or minor obstacles. - Keep
web_searchout of production for now. The current Kimi API guide says the feature is being updated and is not recommended for near-term production workflows.
These caveats apply to the current pay-as-you-go API contract. Kimi Code membership is a separate managed product: its k3 route currently exposes low/high/max effort, while 256K versus 1M context depends on the membership tier.
CAR Cannot Be Computed Yet
AIHackers cannot compute Cost per Accepted Result (CAR) for K3 until there is an accepted-task field test.
| |
Run this five-task first pass before making K3 a default route:
| Task | Pass condition |
|---|---|
| Repository map | Correct modules, commands, and constraints; no invented files |
| Bug fix | Minimal patch for one real failing test |
| Refactor | 2-4 files changed while preserving local style and behavior |
| Code review | File-grounded findings on a known risky patch |
| Multimodal task | Screenshot or image-grounded UI/code task that uses visual evidence concretely |
Record exact model ID, reasoning_effort, context size, token usage, cache hit/miss behavior, retries, latency, accepted result, and review minutes. If K3 produces long reasoning traces, the $15 output price matters more than the headline input price.
Open-Weight Release Boundary
Moonshot’s repository now provides the license, model configuration, tokenizer, and serving guidance. The release resolves the promised-weight question but not the operating-cost or accepted-result question. Before production self-hosting, verify:
- License terms, commercial-use terms, acceptable-use policy, and any geographic restrictions.
- Checkpoint size, shard format, tokenizer, and exact architecture/config files.
- Whether vLLM, SGLang, KTransformers, or other engines support KDA and the K3 checkpoint without vendor-only patches.
- Realistic self-hosting requirements for the 64+ accelerator guidance.
- Independent reproduction of headline benchmark results.
- Whether Kimi API docs add low/high reasoning effort modes or pricing changes after launch.
The released weights do not make Moonshot’s managed API free, and they do not establish CAR. Hardware, serving labor, review time, and accepted output still belong in the measurement.
Related links
- /models/ - current model status guide
- /models/kimi-k2.7-code/ - cheaper Kimi coding API lane
- /tools/kimi-code/ - Kimi Code membership and model configuration
- /value/kimi-access/ - Kimi access routes and migration notes
- /value/smart-spend/ - paid-stack routing strategy
- /compare/models/mid-range/ - production spend-band comparison
- /posts/how-to-read-ai-benchmarks/ - benchmark and CAR method
Sources
- Kimi global model and pricing surface
- Kimi K3 launch blog (Archive)
- Kimi K3 quickstart (Archive)
- Kimi model list (Archive)
- Kimi K3 pricing (Archive)
- Kimi K3 weights and license (Archive)
- Kimi Code model configuration (fresh archive pending after exact/protocol/wildcard retries and an HTTP 520 save response on July 18)
- Artificial Analysis: Kimi K3 (Archive)
- Artificial Analysis: Claude Opus 4.8 (fresh v4.1 archive pending after an HTTP 520 save response on July 18)
Last verified: August 1, 2026. Kimi K3 prices, model-list status, reasoning modes, independent scores, and serving support can change quickly.