By ByteBot
Moonshot AI’s Kimi K3 — a 2.8-trillion-parameter model that topped the Frontend Code Arena leaderboard on day one — releases its full weights on July 27. Seven days. If you have been waiting for a frontier-class open model you can run privately, deploy in an air-gapped environment, or use to escape $15-per-million-token output pricing from closed providers, this is the release you marked on your calendar. Here is what to do before the download button appears.
Stop. Do the Hardware Math First.
Before anything else: K3 is not a consumer GPU story. Its MXFP4 weights require approximately 1.4 TB of storage — a 4x reduction from the 5.6 TB FP16 equivalent, but still a number that eliminates most desks. Moonshot’s own deployment guidance calls for 64 or more accelerators in supernode configurations. A realistic minimum is an eight-node cluster with eight 80 GB GPUs each, giving you 5.12 TB aggregate VRAM with room for KV cache and activations.
The MXFP4 format is natively supported on NVIDIA Blackwell (H200, B100) and AMD MI400. If you are running H100s, plan for a format conversion step — tooling will emerge around the weight release but is not guaranteed on day one. A practical threshold: if you do not have at least four H100s available, the API is the correct call on July 27, not a download.
The API Is Live Today — Three Things to Fix
You do not need to wait for weights to start building with K3. The API has been available since July 16 at api.moonshot.ai/v1 with model ID kimi-k3. It uses OpenAI’s Chat Completions schema, so pointing an existing client at the new base URL with a Moonshot API key is enough to get started. That said, K3 has quirks that will break your code if you carry assumptions from other models.
First: reasoning_effort accepts only max. There is no lighter reasoning mode. K3 always thinks before answering, and those thinking tokens bill at the output rate ($15 per million). If your cost model assumed K3 would behave like a fast-response model, recalculate.
Second: Do not pass temperature, top_p, or seed. K3 does not accept sampling parameters — they are fixed internally. Passing them will either error out or be silently ignored depending on client version.
Third: Preserve complete assistant messages across tool call chains, including reasoning content. K3’s architecture requires the full context to remain intact for multi-step tool use. Truncating the reasoning block between turns produces degraded results in agentic workflows.
from openai import OpenAI
client = OpenAI(
base_url="https://api.moonshot.ai/v1",
api_key="YOUR_MOONSHOT_API_KEY"
)
response = client.chat.completions.create(
model="kimi-k3",
# No temperature, top_p, or seed — K3 rejects them
messages=[
{"role": "user", "content": "Review this function for security issues:\n\ndef login(user, pw):\n query = f'SELECT * FROM users WHERE name={user} AND pass={pw}'\n return db.execute(query)"}
]
)
print(response.choices[0].message.content)OpenRouter also carries K3 at identical pricing (moonshotai/kimi-k3) if you want provider redundancy or prefer a single API key across models.
What to Prepare for July 27
Moonshot committed to releasing a Kimi Delta Attention (KDA) prefill cache implementation to vLLM alongside the weights. If you plan to self-host, pull the latest vLLM release on or after July 27 before pinning production images — KDA support is the critical dependency for K3 to run efficiently.
The recommended launch configuration uses tensor parallelism across eight GPUs with a 128k context ceiling on day one:
vllm serve moonshotai/kimi-k3 \
--tensor-parallel-size 8 \
--max-model-len 131072The full 1M context window is a cloud serving feature, not a local one — at least initially. Treat 128k to 256k as the realistic first milestone for local inference. SGLang and TensorRT-LLM ports will follow, but their timing is community-driven and uncertain.
Which Path Is Right for Your Team
Three honest options, without the vendor spin:
- Cluster (64+ H100s): You get the full model, the full context window via self-host eventually, and the compliance guarantee of on-premise inference. Enterprise or hyperscaler territory.
- Hybrid routing: Run K2.7 Code locally for latency-sensitive, repetitive tasks. Route complex reasoning and long-context work to the K3 API. This is the highest-ROI setup for most engineering teams — you keep local speed where it matters and access frontier capability where it counts.
- Pure API: Use
api.moonshot.ai/v1or OpenRouter. $3 input, $15 output per million tokens. At roughly one-third the price of Claude Fable 5 and GPT-5.6 Sol for comparable coding performance, this is the default-correct choice for individuals and small teams.
What the Benchmarks Actually Say
K3 debuted at #1 on LMArena’s Frontend Code Arena with a 1,679 Elo score, ahead of Claude Fable 5 (1,631) and GPT-5.6 Sol (1,618). On SWE-bench Verified it posted 76.8%; Terminal-Bench 2.1 came in at 88.3%. These are not synthetic toy benchmarks — Frontend Code Arena uses real human preference votes, and SWE-bench tests actual repository patch generation.
On overall intelligence rankings, K3 sits at 57.11 on the Artificial Analysis Intelligence Index — on par with Claude Opus 4.8 and GPT-5.5, but behind Claude Fable 5 and GPT-5.6 Sol. The honest read: K3 is a frontier-adjacent model that competes at the top specifically on coding tasks, and does so as an open model anyone will be able to download. That combination has not existed before at this quality level.
Mark July 27: Three Things to Check When Weights Land
When the release drops, do not just download. Verify three things before committing to a deployment path:
- vLLM KDA merge status. Check the vLLM changelog that day. The KDA implementation should land in the same window; if it has not, hold off on production deployment.
- Context window per your hardware tier. Community reports will surface quickly. Actual usable context on your GPU count may differ from the theoretical maximum.
- MXFP4 conversion tool availability. If you are on H100s rather than Blackwell, you need a weight conversion step. Check the Moonshot HuggingFace repo and vLLM issues for the recommended path.
K3’s prior models (K2.6, K2.7 Code) cleared 1.4 million and 781k HuggingFace downloads respectively in their first months. Expect demand to spike immediately — plan your download window accordingly.
