Models

Kimi K3.1 Pro is open weight, not open hardware

Moonshot AI published the full weights for Kimi K3.1 Pro on November 5, keeping a promise it made at the October 24 launch. The download is real. What you can do with it depends on hardware almost nobody owns.

Daniel HarrisDaniel Harris
Heat: 1,150
Indigo lattice forming a hexagon to represent an open-weight model release

Moonshot AI published the full weights for Kimi K3.1 Pro on November 5, keeping a promise it made at the October 24 launch. The download is real. What you can do with it depends on hardware almost nobody owns.

What Moonshot shipped with Kimi K3.1 Pro

The release landed on Hugging Face, and it covers the whole model rather than a distilled cut. That distinction matters. A distilled release runs on smaller machines but gives you a weaker model, and this isn't that.

The numbers: 2.8 trillion total parameters, roughly 104 billion active per token, a mixture-of-experts design, a context window of about 1,048,576 tokens, and native image understanding. Moonshot delivered all of it on the date it said it would, two weeks after the launch event.

Prompting a model with a million tokens is a different habit from prompting one with 128,000, and very few people have the hardware to try it at home.

Kimi K3.1 Pro is a post-trained K3, not a new base

Same parameter count as Kimi K3. Different training after the fact. Moonshot describes K3.1 Pro as a post-trained version of K3, with much larger reinforcement-learning post-training across hundreds of downstream domains.

Reinforcement learning, in this setting, means training the model on its own attempts and rewarding the ones that work. The base capability stays where it was. The behaviour changes. Expect better instruction following and more recovery from its own mistakes.

That's the pattern most labs have settled into this year. Base models are expensive and slow to build, so the visible progress shows up in the post-training pass instead.

Gated sparse attention inside Kimi K3.1 Pro, explained simply

Long context is expensive because attention compares every token against every other token. Moonshot replaced some full-attention layers with gated sparse attention, built from two parts: KDA linear attention and DSA sparse attention. The aim is to cut the cost of working through a very long document.

For a reader, the payoff is a 1M context model that doesn't collapse under its own weight. For an inference provider, the payoff is a bill that doesn't scale quite so brutally with sequence length.

Why Kimi K3.1 Pro's 104 billion active parameters set the serving cost

A mixture-of-experts model routes each token through a small slice of its total parameters. Compute per token tracks the active count, not the 2.8 trillion headline. That's why the architecture exists, and it's why a model this size can serve at a sane speed when a dense model of the same total would crawl.

Memory is the other half of the bill, and memory doesn't care about routing. Every parameter has to sit somewhere. A provider serving a 1M context model keeps the whole thing resident on expensive memory, so a throughput number that looks fine on a short prompt can fall apart on a long one.

The 1M context model comes with a 5.6 TB bill

At BF16 the weights run about 5.6 TB. At INT4 they drop to roughly 1.4 TB. Either figure puts a consumer GPU out of the running, and a workstation with a few cards doesn't get close.

Huawei says Ascend hardware supports day-one inference through vLLM-Ascend and SGLang. That's the practical path for anyone who isn't renting time inside a large cloud.

Agent Arena ranks Kimi K3.1 Pro first among open-weight models

On Arena's Agent Arena, Kimi K3.1 Pro (Max) sits first among open-weight models and 13th overall. The net improvement is +5.1% against the average, where Kimi K3 (Max) managed +0.9%, measured over more than 17,500 real long-horizon agent sessions.

The sub-scores show where the gain came from. Bash Reliability rose 7.9%, after K3 sat at -4.5%. Praise versus Complaint gained 7.7%, Steerability 6.5%, and Confirmed Success 3.9%. Those are the behaviours that decide whether an agent finishes a job or wanders off halfway through it.

Reliability on shell commands is an odd thing to lead with until you watch an agent work. Most long tasks die on a failed command, not on a bad plan.

Terminal-Bench 4.0 puts Kimi K3.1 Pro behind Claude Opus 5.5

Moonshot's own table lists Kimi K3.1 Pro at 65.7% on Terminal-Bench 4.0, a test of command-line work. Claude Opus 5.5 sits at 66.4% and GPT-6 Astra at 58.2% in the same table. A vendor table isn't a neutral testing ground, so treat the ordering as a claim until someone reproduces it.

That's the honest read, and it's also the useful one. A gap of less than a point against a closed frontier model is close enough to matter for anyone budgeting inference across thousands of agent runs a month.

Who the Kimi K3.1 Pro open-weight release is for

So who benefits from a release most people can't run? Anyone who sells inference, mostly.

Worth saying plainly.

Not laptop users. The Kimi K3.1 Pro download is aimed at labs that want to fine-tune, inference providers that plan to host it behind an API, and companies that need a model on their own metal for legal or latency reasons.

Seen from the outside, the Moonshot AI open weights release is a statement about where the company wants to sit in the market. Open weights get a model into more hands than a hosted API does, and they make a small lab a default choice for teams that want to keep data in-house.

The model card lists the licence terms. Anyone planning a commercial deployment should read them before pulling 5.6 TB through their pipes.

What the Kimi K3.1 Pro download changes

An open-weight model at this size moves the competitive question. It's no longer whether an open release can match the frontier in a given week. It's whether anyone outside a hyperscaler can afford to serve the result.

Moonshot answered the first part by publishing. The second part belongs to whoever pays the electricity bill.

Share This Story

Sources

Related AI News

A terminal window with a rising reward curve and a 27B badge beside it
Models

RadixArk Fine-Tuned a 27B Model on 1,200 Tasks and It Beat the Original Everywhere

RadixArk took an existing open model, spent 18 days and about 1,200 hand-built tasks on it, and changed nothing about its shape. The gains are real, and the absolute scores are still low.

Heat: 1,330
A laptop outline with three stacked layer sheets passing through a loop arrow inside it
Models

Inkling 2 Mini Fits on a Laptop by Running the Same Layers Three Times

Thinking Machines Lab built a model that thinks longer by running the same block of layers three times. The trick buys reasoning without extra memory, and the bill comes due somewhere else.

Heat: 1,710
A long teal audio waveform with a single orange accent bar on a dark field
Models

A 0.5B Speech Model Just Matched a 30B One on Real Phone Calls

Reson8 shipped a half-billion-parameter transcription model that gets within a tenth of a point of a 30B rival on noisy phone calls. The interesting part is what that does to a per-minute bill.

Heat: 560
Three neon rings orbiting a small glowing core on a deep blue field
Models

Gemini 4 Neon Lands in the App, and the Model Picker Loses Two Names

Google's fourth Gemini model in a month slots into the middle of the lineup, priced to be used every day. The app now asks subscribers to choose by effort, and two familiar names are gone.

Heat: 1,500