Moonshot AI shipped Kimi K3 yesterday: 2.8 trillion parameters, a 1-million-token context window, and benchmark scores within striking distance of GPT-5.6 Sol — priced at $3/$15 per million tokens, the same as Claude Sonnet 4.x. By parameter count, it is the largest open-weight model ever released. The API is live now. The full weights drop on July 27.
The Pricing Is the Story #
Kimi K3 is not cheap. At $3/$15 per million input/output tokens, it costs three times more than its predecessor K2.6 ($0.95/$4). Moonshot is done competing on price alone, and that tells you something about where the open-weight market is heading.
But context matters. K3 at $3/$15 still undercuts GPT-5.6 Sol ($5/$30) by roughly 40% on input and 50% on output, while offering open weights that Sol will never have. The question for your stack is not whether K3 is expensive — it’s whether frontier-class open-weight performance at Sonnet-tier pricing changes your build calculus. For a lot of teams, it does.
| Model | Input / 1M | Output / 1M | Open Weights? |
|---|---|---|---|
| Kimi K3 | $3.00 | $15.00 | July 27 |
| GPT-5.6 Sol | $5.00 | $30.00 | No |
| GPT-5.6 Terra | $2.50 | $15.00 | No |
| Claude Fable 5 | $10.00 | $50.00 | No |
| DeepSeek V4 | <$1.00 | <$5.00 | Yes (MIT) |
What the Benchmarks Actually Say #
The AI news framing that will follow you everywhere: Opus 4.8-class at Sonnet 5 pricing. That’s roughly right. Artificial Analysis puts K3 at 57.1 on its Intelligence Index, against GPT-5.6 Sol’s 58.9 and Fable 5’s 59.9. Two points behind the closed frontier, at half the price, with weights coming.
Moonshot’s internal benchmark table shows DeepSWE at 67.5, Terminal-Bench 2.1 at 88.3, and FrontierSWE at 81.2. On agentic tasks, K3 scores 91.2 versus GPT-5.6 Terra’s 87.4. Take those numbers seriously, but not uncritically — independent SWE-bench Verified and LiveCodeBench numbers have not landed yet. Moonshot grading Moonshot is not the same as third-party evaluation.
One useful data point that isn’t from Moonshot: K3 is currently leading Arena.ai’s Frontend Code arena, including above Fable 5. That’s community voting, not a controlled benchmark, but it correlates with real-world coding quality.
The Architecture That Makes 1M Context Usable #
K3 is built on Kimi Delta Attention (KDA), a hybrid linear attention mechanism that Moonshot describes as a refinement of Gated DeltaNet with finer-grained memory gating. In practice: three of every four layers use cheap linear attention, with one full-attention layer at every fourth position. The result is 6.3x faster decoding at million-token context lengths and a 75% reduction in KV-cache memory — which means 1M-context runs that are actually fast enough to use in production, not just benchmarks.
The other architectural piece, Attention Residuals (AttnRes), selectively retrieves representations across model depth. Moonshot hasn’t published a paper on it yet — the name first appears in the K3 docs themselves. The full technical report is expected alongside the weight release on July 27. K3 achieves roughly 2.5x the scaling efficiency of K2, which is the practical payoff from these architectural changes.
Two Variants: Pick the Right One #
K3 Max handles the standard agentic loop: coding tasks, research, multi-turn document work, long-horizon planning sessions. This is the one most developers will reach for.
K3 Swarm Max is for parallel multi-agent workloads — think large-repo analysis, batch file processing, or the kind of 100-plus coordinated sub-agent runs that Moonshot’s Agent Swarm system enables. If you are orchestrating agents at scale rather than running a single long session, Swarm Max is the right tool.
Using K3 Now #
The Kimi K3 API is OpenAI-compatible. Drop in your Moonshot API key and swap the base URL:
from openai import OpenAI
client = OpenAI(
api_key="YOUR_MOONSHOT_API_KEY",
base_url="https://api.moonshot.ai/v1",
)
response = client.chat.completions.create(
model="kimi-k3",
messages=[{"role": "user", "content": "Your prompt here"}]
)
print(response.choices[0].message.content)
One thing to note: K3’s sampling parameters (temperature, top_p, penalty settings) are fixed server-side. Do not include them in your requests.
When the weights land on July 27, an early Q4 GGUF quantization is already on HuggingFace. For local deployment via llama.cpp, expect to need a 24GB GPU and at least 240GB of disk for the quantized version, with MoE layers offloaded to system RAM. Ollama support requires a manual patch to LLAMA_MAX_EXPERTS (256 to 384) and a recompile until the next Ollama release catches up.
What This Means for Your Stack #
The short version: open-weight and closed-model frontier performance are now within a few benchmark points of each other, and the price gap is compressing. If you are currently paying for GPT-5.6 Sol or Claude Opus 4.8 for agent workloads, K3 deserves a controlled pilot before your next invoice. If you are running DeepSeek V4 for cost reasons and need vision or better agent coordination, K3 closes the capability gap at a price that is still workable.
The signal underneath the benchmarks: Moonshot tripling its prices is not a market retreat. It’s a bet that the open-weight tier has graduated from “cheap alternative” to “legitimate frontier option.” The July 27 weight release will tell us whether the community agrees.