Kimi K3.1 and K3.1 Pro Explained: Specs, Pricing, Open Weights
Kimi K3.1 Pro from Moonshot AI: 3T-class, 1M-token context, $2/$10 per million tokens, and weights now open on Hugging Face. Specs, API tips, comparisons.

Kimi K3.1 Pro landed on October 24, 2026, and spent that day at the top of Hacker News with 504 points. Moonshot AI positions it as its most capable 3T-class open model. Update, November 5: the asterisk is gone — the weights are now public on Hugging Face. The API remains the practical way to run it today, and this post now covers what the open release actually gives you.
If you searched for "Kimi K3.1" expecting a base model plus a Pro tier, adjust expectations. The K3 series pricing page lists kimi-k3.1-pro and kimi-k3, and no standalone K3.1. In practice "Kimi K3.1", "Kimi 3.1 Pro" and "K3.1 Pro" all point at the same thing — the model this post covers.
Below: the spec sheet, what the Highspeed variant changes, how to call it today (including the caching and Batch rules that decide what you actually pay), what the open-weights release actually ships, and how it sits next to DeepSeek V4.1 and GLM-5.3.
What "Kimi K3.1" actually refers to
Kimi is the model family from Moonshot AI, and K3.1 is the current generation. Two things ship under that label:
- Kimi K3.1 Pro — the flagship, API name
kimi-k3.1-pro. - Kimi K3.1 Pro Highspeed — the same model, served roughly 2-3x faster, at about 2x the price.
There is no cheaper "K3.1 base" sitting under the Pro. The K3 series page lists kimi-k3.1-pro and the previous-generation kimi-k3; if you want the older model, that row is the one to look at, not a K3.1 variant that does not exist.
The other thing happening in the lineup is the K2 retirement. New K2 registrations closed on October 31, 2026, and the series is scheduled for full retirement by the end of December 2026. K2 topped out at a 262,144-token context window; K3.1 Pro quadruples it. If you are still building on a K2 model, you have roughly two months to move.
Kimi K3.1 Pro specs at a glance
| Spec | Kimi K3.1 Pro |
|---|---|
| API model name | kimi-k3.1-pro |
| Released | October 24, 2026 |
| Context window | 1,048,576 tokens (1M) |
| Input price | $2 per million tokens |
| Output price | $10 per million tokens |
| Batch API | 60% of standard ($1.20 / $6 per million) |
| Context caching | On by default, 5-minute TTL |
| Terminal-Bench 4.0 | 65.7 |
| Highspeed variant | Roughly 2-3x faster, about 2x the price, same capability |
| Open weights | Released — moonshotai/Kimi-K3.1-Pro on Hugging Face |
| Positioning | "Most capable 3T-class open model" (Moonshot) |
Two notes on reading that table. First, "3T-class" is the label Moonshot itself uses — read it as a statement about parameter scale, not a published architecture sheet; we are not going to guess at active-parameter counts the company has not stated in the materials we checked. Second, Terminal-Bench 4.0 measures agents doing real work in a terminal — installing things, fixing builds, chaining commands — which is exactly the workload Moonshot is pitching this model at. A 65.7 is a strong number, but benchmark scores only mean something next to other models run on the same harness, and we do not have same-harness figures for the competitors below, so we will not pretend otherwise.
Standard vs Highspeed
Same weights, different serving tier. Highspeed is roughly 2-3x faster at about twice the price — if the multiplier holds exactly, budget around $4 input and $20 output per million tokens, and confirm the live number on the pricing page before you commit.
When it is worth it: interactive agent loops where a human waits on every turn. A coding agent that makes 40 tool calls per task feels like a different product at 2x speed. When it is not: batch jobs, offline evals, anything asynchronous. Pay for latency only where someone is watching a spinner.
How to call Kimi K3.1 Pro today
The API lives at platform.kimi.ai — the docs moved there from platform.moonshot.ai, so update any old bookmarks. The model string is kimi-k3.1-pro:
{
"model": "kimi-k3.1-pro",
"messages": [
{ "role": "system", "content": "You are a code reviewer." },
{ "role": "user", "content": "Review the diff below and flag risky changes." }
]
}
Three billing mechanics decide what a 1M-context model actually costs you:
Context caching is on by default. Cached prefixes live for 5 minutes. Cache writes are billed by TTL tier, and cache hits pay only the cache-hit price — lower than the $2 standard input rate, with the exact figure on the pricing page. Put the stable material (system prompt, repo map, long documents) at the front of the prompt and the per-turn delta at the end, so each agent step re-reads the same prefix from cache instead of paying full price for it. If your calls land more than 5 minutes apart, assume a cold cache.
Batch runs at 60%. Anything that does not need an answer now — nightly evals, bulk summarization, re-indexing — goes through the Batch API at $1.20 input and $6 output per million tokens. That is the single biggest lever on the bill.
1M context is a ceiling, not a target. At $2 per million input tokens, a full-window call costs about $2 before you count output. Fine for a one-shot "read this entire repository" task; ruinous if an agent does it every turn. Pair the big window with caching, and prune what you send.
If you would rather not wire the API yourself, Moonshot's terminal agent Kimi Code is the first-party path — check which K-series model your install is pointed at before assuming it is K3.1 Pro.
Open weights: what actually shipped
Update (November 5): the weights are no longer scheduled — they are up. The repo is moonshotai/Kimi-K3.1-Pro on Hugging Face: 96 safetensors shards totalling roughly 1.6 TB, ungated, with transformers support and 8-bit compressed-tensors weights. Unsloth mirrored the repo on November 4; community quants will follow.
Two things the model card settles:
- Vision is in. The repo is tagged image-text-to-text — the open weights take image input, which settles the multimodal question this post originally flagged as unverified.
- The license is custom. The Kimi K3.1-Pro License is permissive at its core — use, copy, modify, merge, publish, distribute, sublicense, sell — with two clauses aimed at very large businesses: if you run a model-as-a-service business and your aggregate revenue with affiliates exceeded $20 million over the trailing 12 months, you need a separate agreement with Moonshot before commercial use; and if your product exceeds 100 million monthly active users or $20 million in monthly revenue, "Kimi K3.1-Pro" must be prominently displayed in your UI. Internal use is exempt.
What has not changed is the physics. A ~1.6 TB checkpoint is multi-node territory — auditability is available today, third-party hosting should arrive over the coming weeks, and practical self-hosting stays a job for well-equipped teams.
Kimi K3.1 Pro vs DeepSeek V4.1 and GLM-5.3
The three names people are weighing against each other this quarter.
| Model | Context | Input per M | Output per M |
|---|---|---|---|
| Kimi K3.1 Pro | 1,048,576 | $2.00 | $10.00 |
| Kimi K3.1 Pro Highspeed | 1,048,576 | About 2x standard | About 2x standard |
| DeepSeek V4.1 Flash | 1M | $0.15 off-peak / $0.30 peak | $0.60 off-peak / $1.20 peak |
On list price, DeepSeek Flash is roughly 13x cheaper on input and 17x on output off-peak — before counting Flash's own $0.003 cache-hit rate. DeepSeek's heavier V4-Pro sits at $0.66 input and $1.98 output off-peak, so K3.1 Pro is about 3x V4-Pro on input and 5x on output.
Kimi is not selling on price. Its case is the 1M window plus terminal-agent capability, and Terminal-Bench 4.0's 65.7 is the evidence it offers. If your own evals confirm that lead on your tasks, the price gap is the cost of fewer failed runs; if they do not, Flash's price is hard to argue with. We break down Flash's tiers and the V4-Pro trade-off in our DeepSeek 4.1 Flash post, with hands-on setup in the DeepSeek V4.1 guide. GLM-5.3 is the third name in this cycle and gets its own breakdown.
Who should pick Kimi K3.1 Pro
Pick it if your workload is long-context and agentic: repository-scale refactors, multi-hour terminal sessions, document sets that genuinely exceed 262K tokens (the K2 ceiling). The 1M window plus default caching is built for exactly that, and Terminal-Bench 4.0 is the benchmark Moonshot chose to lead with for a reason.
Skip it for now if you need third-party self-hosting on day one (the weights are out, but hosting providers take weeks to catch up); you are cost-constrained at high volume (DeepSeek Flash is an order of magnitude cheaper); or your pipeline needs image input — the open-weights repo is tagged image-text-to-text, but confirm the API tier you call actually exposes vision before committing, since this post covers text pricing and context only.
If you are choosing between terminal agents rather than raw models, our AI coding tools like Cursor roundup covers Kimi Code alongside Claude Code and the rest.
FAQ
Is there a Kimi K3.1 that is not Pro?
Not on the K3 series pricing page, which lists kimi-k3.1-pro and kimi-k3. Searches for "Kimi K3.1" and "Kimi 3.1 Pro" both resolve to K3.1 Pro in practice. If you want the previous model, kimi-k3 is the one still listed.
How much does Kimi K3.1 Pro cost?
$2 per million input tokens and $10 per million output tokens on the standard tier. Batch requests are billed at 60% of that ($1.20 / $6). Context caching is on by default with a 5-minute TTL, and hits pay the lower cache-hit rate. Highspeed is about 2x the standard price for 2-3x the speed.
When can I download the Kimi K3.1 Pro weights?
Now. The weights are on Hugging Face at moonshotai/Kimi-K3.1-Pro — 96 safetensors shards, roughly 1.6 TB, no download gate. Expect third-party hosting to pick it up over the coming weeks; practical self-hosting of a 3T-class model stays limited to well-equipped teams.
What license do the Kimi K3.1 Pro weights use?
The custom Kimi K3.1-Pro License: permissive for individuals and most companies — use, modify, sell — with a revenue gate. MaaS providers whose aggregate revenue with affiliates exceeded $20 million over the trailing 12 months need a separate Moonshot agreement, and products over 100 million MAU or $20 million monthly revenue must display "Kimi K3.1-Pro" in their UI. Internal use is exempt.
What is the Kimi K3.1 Pro context window?
1,048,576 tokens — 1M — on both the standard and Highspeed tiers. That is 4x the 262,144-token window of the retiring K2 series.
Kimi K3.1 Pro or DeepSeek V4.1 Flash?
Flash is roughly 13x cheaper on input off-peak and is the only DeepSeek model with vision. K3.1 Pro is the long-context terminal-agent play with a 65.7 on Terminal-Bench 4.0. Run both on a slice of your real tasks — the price gap only pays for itself if the extra capability shows up in your evals.


