The number that matters: 2,800,000,000,000
That's the parameter count on Kimi K3, the model Moonshot AI dropped on July 16 and which promptly hit 1,587 points and 937 comments on Hacker News — the kind of engagement that makes every other story on the front page look like a rounding error.
The pitch: the largest open-weight model ever released, going toe-to-toe with GPT-5.6 and Claude on benchmarks that actually resemble work, not trivia. And unlike most "open" model announcements, this one has numbers attached instead of vibes.
Here's what's real, what's marketing, and what you should actually do with it.
What's under the hood
This isn't a bigger version of last year's transformer. Moonshot changed the internals:
- 2.8T total parameters, Stable LatentMoE architecture, 16 active experts out of 896
- Kimi Delta Attention (KDA) + Attention Residuals (AttnRes) — their own attention variant, not vanilla MHA/GQA
- 1,048,576 token context window (~1,573 pages)
- MXFP4 weights, MXFP8 activations — shipped quantized from the start, not bolted on after
- Claimed ~2.5x scaling efficiency over Kimi K2
- Native vision input, text output, max reasoning effort on by default — there's currently no low-effort mode
The MoE ratio is the tell here. 16-of-896 is an extremely sparse activation pattern — you're paying inference cost closer to a much smaller dense model while keeping a knowledge base the size of a small country's worth of text. That's the actual engineering bet: sparsity buys you scale without the GPU bill scaling linearly with it.
The benchmarks, unromanticized
On GDPval-AA v2 (Artificial Analysis's benchmark spanning 44 real occupations across 9 industries):
| Model | Score |
|---|---|
| Claude Fable 5 Max | 1,815 |
| GPT-5.6 Sol Max | 1,747.8 |
| Kimi K3 | 1,687 |
| Claude Opus 4.8 | 1,600 |
Third place, but ahead of Opus 4.8 — which is not nothing.
On AA-Briefcase (long-horizon agentic knowledge work), it does better:
| Model | Score |
|---|---|
| Claude Fable 5 Max | 1,587 |
| Kimi K3 | 1,527 |
| GPT-5.6 Sol Max | 1,495 |
Second place, ahead of GPT-5.6. Artificial Analysis's own Intelligence Index puts it at 57, good for #4 out of 189 tracked models, against a field average of 30.
So: not the best model in the world. Best open model by a wide margin, and competitive with closed frontier models on the benchmarks that measure "can this thing actually do a job" rather than "can this thing pass a trivia set."
Pricing — and the catch inside the pricing
Input (cache miss): $3.00 / 1M tokens
Input (cache hit): $0.30 / 1M tokens
Output: $15.00 / 1M tokens
Moonshot claims >90% cache hit rates on coding workloads, which would make real-world cost closer to the $0.30 number than the $3. That's a meaningful claim if true — most teams running agentic coding loops re-send huge chunks of unchanged context on every turn, and a 90% cache hit rate turns this into one of the cheaper frontier-tier options on the market.
But Artificial Analysis's own eval run tells a less flattering story: 62 tokens/second, ranked #88 out of 189 — slower than the field median of 70 t/s. And in Simon Willison's now-standard "draw an SVG pelican riding a bicycle" test, K3 burned 13,241 reasoning tokens to produce a 3,417 token response, for a task that costs single digits of tokens on a lighter model. Willison also flagged that the prompt itself — 10 tokens on OpenAI's tokenizer — came out to 95 tokens on Kimi's, implying roughly an 85-token hidden system prompt you're paying for on every single call.
Translate that: the sticker price looks great, the cache-hit discount looks great, and then "max reasoning effort, no dial to turn it down" quietly eats the savings on anything that isn't a cache-friendly coding loop.
Calling it
curl https://api.moonshot.ai/v1/chat/completions \
-H "Authorization: Bearer $KIMI_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "kimi-k3",
"messages": [
{"role": "user", "content": "Refactor this function for readability, keep behavior identical."}
]
}'
Available now on kimi.com, Kimi Code, Kimi Work, and the API. Weights land July 27 — until then, despite the "open" branding everywhere in the coverage, Artificial Analysis correctly classifies it as proprietary. You cannot self-host this yet. Anyone writing "download Kimi K3 today" this week is wrong.
The actual takeaway
Ignore the "world's largest open model" framing — that's a vanity metric, and raw parameter count stopped being a proxy for quality two years ago. The number that should hold your attention is 1,527 on AA-Briefcase, beating GPT-5.6 Sol on long-horizon agentic work, from a lab that's going to publish its weights in public in under two weeks.
That's the actual story: the gap between "best closed model" and "best model you can eventually run yourself" just got smaller again, on a benchmark that measures whether a model can hold a job, not whether it can ace a leaderboard. Whether it's worth switching your agent stack over depends entirely on whether your workload looks like Moonshot's cache-hit benchmark (cheap) or Willison's pelican test (expensive) — and right now there's no dial to move you between the two.
Watch July 27. The pricing story and the benchmark story are both true. The self-hosting story hasn't happened yet.
Top comments (0)