Kimi K3.1 Pro Open Weights: Self-Host or Use the API?

Moonshot AI put the full weights for Kimi K3.1 Pro on Hugging Face on November 4, 2026: 2.8 trillion total parameters, 104 billion active per token, and a 1.56 TB download. The honest answer to "should I self-host it" is no, unless you're already burning more than roughly $35,000 a month on the Moonshot API. For almost everyone else, the API at $2 per million input tokens and $10 per million output tokens is the only setup that makes financial sense.

I went through the license, the hardware math, and the actual benchmark numbers (not just Moonshot's own slide deck) to figure out where the break-even point actually sits. Short version: it's much higher than the marketing implies.

What Is Kimi K3.1 Pro, Exactly?

Kimi K3.1 Pro is Moonshot AI's flagship open-weight model, built on a Mixture-of-Experts architecture with 93 layers (a mix of 69 KDA layers and 24 Gated DeepSeek Sparse Attention layers, if you care about the internals). Only 104 billion of the 2.8 trillion total parameters activate on any given token, which is the only reason this thing is runnable at all.

The specs that matter:

• 2.8 trillion total parameters, 104 billion active (MoE)
• 1.56 TB of weights, shipped as MXFP4 weights with MXFP8 activations
• 1,048,576 token context window
• Multimodal — takes text and images
• Published on Hugging Face as moonshotai/Kimi-K3.1-Pro, with deployment recipes for Transformers, vLLM, and SGLang

Moonshot is pitching it hard at coding agents, full-project game work in Unreal, Unity, and Godot, parametric CAD modeling, and finance tasks. That's a wider net than the usual "better at SWE-bench" pitch, and it tracks with where Kimi K3 already sat against Claude on frontend work.

Can You Actually Self-Host This Thing?

Technically, yes. Practically, you need a server most companies don't own.

The 1.56 TB weight size means you need somewhere north of 1,680 GB of VRAM just to load the model, before you account for KV cache. Here's what that rules in and out:

So the entry ticket is eight of Nvidia's newest B300-class chips (or AMD's MI355X), not the H100s or H200s most self-hosters already own. That's the detail a lot of the hype coverage skips.

What Does Self-Hosting Actually Cost Per Month?

Depends entirely on how hard you drive the GPUs, and this is where the "self-hosting is cheaper" narrative falls apart for most workloads.

Running the math from actual deployment cost breakdowns:

• Renting an 8x B300 node 24/7 from an overseas cloud provider runs $34,500 to $46,700 a month
• A real-world Modal billing example came in at $41,457/month
• Buying the hardware outright, amortized with electricity, lands around $27,300 to $28,000/month

Now compare that to what you get per million output tokens at different utilization levels:

That breakeven only shows up around 6.5 to 10.4 billion input tokens and 1.3 to 2 billion output tokens a month, sustained. If your GPU cluster sits idle half the day waiting for requests, you're paying for a Ferrari to idle in traffic. One HN commenter on the original announcement thread described running it on a 4x AMD MI300A box (roughly $110k) for 30-40 analysts and coming out ahead of their Claude API bill, which checks out, because that's a team generating constant, predictable load, not bursty requests.

What's Actually in the License?

"Open weights" doesn't mean unrestricted here, and this trips people up every time a Chinese lab ships a model.

The Kimi K3.1 Pro license reads like MIT at first glance, but it adds two commercial conditions:

1. If you're reselling model access as a hosted service and you (plus affiliates) cross $20 million in revenue in any trailing 12-month period, you need a separate agreement with Moonshot.
2. If your product hits 100 million+ monthly active users, you have to display attribution.

Internal use, fine-tuning for your own product, and running it inside your company are all unrestricted. This only bites companies big enough to resell K3.1 Pro as their own API product. If that's not your business model, the license is a non-issue.

Is It Actually Good at Coding?

Mixed, and depends heavily on whose benchmark you trust.

On Moonshot's own SWE-Marathon v1.1 (20 multi-hour engineering tasks, no internet access), K3.1 Pro scored 51.7% against Claude Opus 5.5's 46.3% and GPT-6 Astra's 42.5%. On Terminal-Bench 4.0, it's basically tied with Opus 5.5 (65.7% vs. 66.4%). FourWeekMBA's independent read of the full benchmark sheet put it ahead of Opus 5.5 on 16 of 37 tests total — not a sweep, but not nothing either.

Then there's KingBench, a third-party hands-on coding eval, where K3.1 Pro landed 6th out of 11 models at 53.7%. And on Moonshot's own ProgramBench chart, it scored 31.5 against Claude's 65.0, a gap nobody's spinning as a tie.

The uncomfortable detail from that HN thread: someone ran an analysis showing the model "names hidden requirements in 65% of answers" when probed about its own instructions, which smells like test-set contamination rather than genuine reasoning gains. Take the benchmark wins with a grain of salt, same as you should with the Kimi K3.1 Pro vs. Claude Opus comparisons floating around right now — nearly all of the favorable numbers come from charts Moonshot designed itself.

So, Self-Host or API?

Use the API unless you have a very specific reason not to. At $2/$10 per million tokens, you'd need to burn through billions of tokens a month before a dedicated B300 cluster even breaks even, and that's before counting the engineers who have to babysit a 1.56 TB model's cold starts (about 27 minutes to load and compile, per the deployment guides).

Self-hosting only makes sense if:

• You have a compliance reason data can't leave your infrastructure
• You're running sustained, predictable load above roughly 1,500 tokens/second around the clock
• You already own or can amortize B300/MI355X-class hardware for other workloads

Everyone else: hit the API, see if the coding performance holds up on your actual codebase (not Moonshot's benchmark suite), and revisit self-hosting only once your bill gets uncomfortable. For most teams evaluating open-weight alternatives right now, it's worth putting K3.1 Pro next to DeepSeek V4.1 Flash too, since that's priced to compete directly on the same self-host-vs-API math.

FAQ

How many parameters does Kimi K3.1 Pro actually use per request?

Only 104 billion of its 2.8 trillion total parameters activate per token, thanks to its Mixture-of-Experts design. The full 2.8T figure is what's stored on disk; the active figure is what determines inference speed and cost.

How big is the Kimi K3.1 Pro download?

1.56 TB, spread across more than 100 safetensors shard files on Hugging Face under moonshotai/Kimi-K3.1-Pro. That size alone rules out a single consumer GPU or even most workstation setups.

What GPUs do I need to self-host Kimi K3.1 Pro?

At minimum, eight Nvidia B300 or GB300 GPUs, or eight AMD MI355X GPUs, for roughly 1.68 TB of combined VRAM. Eight H200s (1.13 TB) and eight B200s (1.44 TB) both fall short.

Is Kimi K3.1 Pro free to use commercially?

Internal use is unrestricted. You only need a separate agreement with Moonshot if you're reselling hosted access and cross $20 million in trailing 12-month revenue, or displaying attribution if your product passes 100 million monthly active users.

Is Kimi K3.1 Pro better than Claude Opus 5.5 at coding?

It's close on some benchmarks and far behind on others. It edges out Opus 5.5 on SWE-Marathon and roughly ties it on Terminal-Bench 4.0, but trails badly on ProgramBench and lands 6th of 11 on the independent KingBench eval. Treat any single benchmark claim skeptically.

What does the Moonshot API cost for Kimi K3.1 Pro?

$2 per million input tokens, $10 per million output tokens, and $0.20 per million cached input tokens. A High Speed variant costs double, at $4/$20.

At what point does self-hosting Kimi K3.1 Pro become cheaper than the API?

Roughly 1,500 tokens/second of sustained throughput, which translates to billions of tokens a month. Below that utilization level, self-hosting costs anywhere from 1.6x to 19x more per output token than just paying Moonshot's API price.

Watch and Read More