What is Kimi K3.1 Pro?

An open-weight, multimodal agentic model from Moonshot AI, post-trained from Kimi K3. It is a mixture-of-experts model with 2.8 trillion total parameters, 104 billion active, and a 1,048,576-token context window.

When were the Kimi K3.1 Pro weights released?

Moonshot posted them on Hugging Face on November 4, 2026. The model had been available through the Kimi API since October 23, 2026.

How much does Kimi K3.1 Pro cost through the API?

$2 per million input tokens and $10 per million output tokens on the standard tier. A highspeed version is priced at twice that, except for cached input, and is described as two to three times faster.

Can I run Kimi K3.1 Pro locally?

Not on consumer hardware. The weights are reported at about 1.56 TB, and the published vLLM recipe is reported to need an eight-GPU baseline. OrcaRouter, which sells access to the model, estimates $550,000 to $650,000 per node.

Is Kimi K3.1 Pro free for commercial use?

It uses a custom license. Companies offering third parties model access must sign a separate agreement if group revenue exceeds $20 million. Internal use and end-user products with the model embedded in specific features are carved out from that requirement, and very large products must display the model name.

Kimi K3.1 Pro Weights Drop: 2.8T Params Open Source 🌙

In partnership with

Good morning, AI builders. Moonshot just put the weights of its 2.8-trillion-parameter Kimi K3.1 Pro on Hugging Face. The catch: a 1.56 TB download and a license line drawn at $20M in revenue.

If you've been waiting for a top-tier open model to host yourself, the question has shifted from access to whether the hardware bill makes sense.

Welcome to the 14,155 builders reading today. Forwarded this? Join here.

The 30-second brief:

Kimi K3.1 Pro: 2.8T parameters, 1.56 TB of weights, $2/$10 by API
Meta buys Conductor: its cloud product shuts down on Nov 20, so export now
Data center votes: Ohio restricted them in at least 14 municipalities
Build with AI: test Kimi K3.1 Pro on your own code with a $1 top-up

MOONSHOT AI

Moonshot releases the weights for its 2.8T-parameter Kimi K3.1 Pro

Bar chart of Moonshot-reported scores, Kimi K3.1 Pro vs Claude Opus 5.5: SWE Marathon 51.7 vs 46.3, BrowseComp 92.1 vs 90.8, HLE-Full with tools 70.0 vs 67.7, Terminal-Bench 4.0 65.7 vs 66.4, OSWorld V2.1 79.6 vs 81.8, HLE-Diamond 39.4 vs 60.9

Source: Moonshot AI model card. Self-reported scores.

The scoop: Moonshot posted the Kimi K3.1 Pro weights on Hugging Face on Nov 4, as promised when the API launched in October. It's a 2.8T-parameter mixture-of-experts model with 104B parameters active and a 1M-token context window.

Key details:

The rivalry: Moonshot's own card has it ahead of Claude Opus 5.5 on SWE Marathon (51.7 vs 46.3) and far behind on HLE-Diamond (39.4 vs 60.9).
Independent read: Artificial Analysis lists it at 46 on Index v5.0.1, 22nd of 62, at about 32 tokens per second. It was reported at 55 at API launch, so check the index version before comparing.
Running it: the API costs $2 in and $10 out per million tokens. Self-hosting starts around eight top-end GPUs, which OrcaRouter, a reseller of the model, estimates at $550K-$650K a node.
The catch: the custom license requires a separate deal if you sell model access and your group's revenue tops $20M. Internal use and features embedded in end-user products are carved out.

Our take: For most teams the API beats buying a node, so the weights matter mainly if you have data-residency needs or heavy internal workloads. Moonshot's scores are its own, and the independent ranking is much cooler, so run your hardest real tasks before you switch anything.

Your Next Side Hustle is Already on This List

Ready to stop scrolling job boards and start building something of your own?

The Hustle's free "Side Hustle Ideas Database" uncovers 100 unconventional, actually-actionable ways to earn on the side.

Inside you'll find:

A curated list of 100 real side hustle ideas spanning e-commerce, services, content, and weird corners of the internet most people never think to look
Opportunities at every level, whether you have 5 hours a week or 50, zero capital or a small runway
Offbeat, under-the-radar ideas you won't find in generic "how to make money" listicles

Subscribe for free today and get instant access. Your next income stream is one click away.

Get The Free Guide

META

Meta buys Conductor, and its cloud shuts down on Nov 20

Source: Conductor announcement.

The scoop: Meta has acquired Conductor, the Mac app for running several coding agents in parallel, with the team joining Meta Superintelligence Labs. Axios broke the news, and the price wasn't disclosed.

Key details:

Do this now: if you use Conductor Cloud, right-click a workspace and choose Transfer to local, or use Settings, Account, Export data.
What it was: a harness that ran Claude Code, Codex and others in isolated copies of a codebase, backed by a $22M Series A in March.
Same day: OpenAI agreed to buy Pierre, whose Git and diff tools already power Codex, pending regulatory approval.

Our take: The layer that manages many agents is getting bought by the labs, and a neutral tool that ran everyone's models just lost its neutrality. Keep your agent workflows in plain Git worktrees and scripts, so the next acquisition costs you an afternoon rather than a migration.

BUILD WITH AI

Test Kimi K3.1 Pro on your own code with a $1 top-up

What you'll make: a short Python script that sends a real function from your codebase to Kimi K3.1 Pro, so you can judge it on your work rather than its benchmarks.

Step by step:

Create a key in the Kimi API console, top up at least $1 to unlock the model, and set it as MOONSHOT_API_KEY.
Run pip install --upgrade 'openai>=1.0', then create an OpenAI client with base_url="https://api.moonshot.ai/v1" and api_key=os.getenv("MOONSHOT_API_KEY").
Call client.chat.completions.create with model="kimi-k3.1-pro", reasoning_effort="low" and this prompt: "Find the three most likely bugs in this function, then give me a fixed version," followed by your code.
Rerun it at high and max, and compare the answers, wait time and token counts. Leave out temperature and top_p, since they're fixed.
If it passes, try kimi-k3.1-pro-highspeed in your real workflow. For multi-turn or tool calls, pass back the complete assistant message, not just its content.

Pro move: Caching is automatic, and cached input is 90% cheaper. If your requests are more than five minutes apart, set prompt_cache_options.ttl to 1h.

Open the Kimi quickstart

DATA CENTERS

Ohio voters restrict data centers in at least 14 municipalities

Bar chart of NBC News/SSRS exit polls, share calling each issue decisive or important: government corruption 90%, cost of living 89%, war in Iran 74%, immigration enforcement 73%, AI and data centers 56%

Source: NBC News/SSRS exit polls, via The Decoder.

The scoop: Ohio voters passed bans or other limits on Tuesday in at least 14 municipalities, per Fortune, with 21 of the 41 data center measures nationwide on the state's ballots.

Nationally: Transformer counts over 30 restrictive measures passed, with results still coming in, and cites a Goldman Sachs warning that moratoria could affect about 45% of planned US capacity growth.
Voters: 56% in exit polls called AI and data centers decisive or important, the lowest of five issues.

Our take: AI wasn't what most voters were deciding on, but local permits can still set the pace of the buildout, so factor siting risk into any capacity plan that runs past 2027.

FROM THE CHANNEL

Catch up on the GPT-7 chatter and Qwen 4 leaks

Watch for a quick rundown of what's being said about GPT-7 and Qwen 4, which is still leaks rather than launches, plus a look at Gemini 4 Carbon.

⚡ Rapid fire

🏢 AMD put a price on Taalas: about $2.9B in cash and stock, including up to $1.4B contingent, per its 10-Q.
💰 Fal raised $1.3B in a Bessemer-led round, doubling its reported valuation to $16B since April.
🔬 Xen credited Anthropic and Claude with finding two VM-escape flaws affecting versions 4.6 onward, with no known mitigation beyond patching.
🏛️ An EU court adviser said Gemini answers can infringe press rights in two situations, in a non-binding opinion in Case C-250/25.
👀 Unconfirmed: Apple cut engineers' daily AI cap from $300 to $200, per Bloomberg, with some allowances possibly falling to $100.

🧰 Tool drop

ChatGPT Images 2.6: turn a prompt or photo into an animated GIF, with frame-level edits.
Pangram Image 1.0: detects AI images down to 256px, with heatmaps showing where the AI content appears.
Tencent Hy4: a new model adding vision and computer operation, available by API through OpenRouter.

What Do You Rate Today's Newsletter?

That's the brief. See you tomorrow at 6am.

The World of AI team