Day 7 · 2026-11-17
35 launches checked147 claims quoted word for word179 user reports linked10 verified fixes linked40 corrections publishedLive 25 days5 of 5 founding sponsor spots left
AI launches, checked against real usersLast re-stamped 17 Nov 2026

All receipts / No. 027

No. 027 · ModelStamped 7 Nov 2026

Day 7 receipt

Kimi K3.1 Pro

Moonshot AI · launched 23 Oct 2026
DeliveredDay 7 verdict

Bottom lineA strong open model if you can wait for it; it's slow.

Promised
A frontier-class 3T open-weight model with 1M context, weights by 4 Nov, near Opus 5.5 on coding.
What happened
Weights landed on time and independent boards rank it near the top. Vendor-harness scores shrink in outside runs, and it's slow.
Our verdict
Kept its word on open weights; the benchmark table flattered it.
  • The real cost. Cost per task stays well below Opus 5.5. Agent Native Engineering
  • Where it wins. Full weights landed on Hugging Face on 4 Nov, as promised. Forum thread
  • Where it falls short. Notably slow: 36.1 minutes per index task in Artificial Analysis runs. Artificial Analysis
  • Still missing. No independent timing of the 2-3x faster Highspeed variant. Artificial Analysis

What people say

“It may be slower but it follows the rules better and makes less stupid mistakes, it takes less time to merge the solution.”
— Hacker News commenter
“K3.1 continuously falls back to its default persona and essentially ignores my system prompt.”
— Hugging Face user
Full evidence 4 claims · 7 user reports · 4 sources

Moonshot AI launched Kimi K3.1 Pro, a 3T-class open-weight model with a 1M-token context window, and a Highspeed variant, with full weights promised by 4 Nov.

Launch post: Introducing Kimi K3.1 Pro · Artificial Analysis and Simon Willison date the API launch 23 Oct 2026; Chinese outlets report 24 Oct (Beijing time). The blog post itself is undated. Added on day 15, after the promised weights date of 4 Nov passed.

Why this stamp

Weights landed on time and it ranks near the top, but vendor-harness scores shrink in independent runs and it is slow.

Moonshot said the full weights would arrive by 4 Nov and they did: the model is on Hugging Face under its own licence, and early adopters are already porting and running it. The 1M context window is confirmed by Artificial Analysis and the model card. The headline benchmark claim is the soft spot: the 65.7% Terminal-Bench figure comes from Moonshot's own Claude Code harness, Artificial Analysis later revised its index and now lists the model at 46, and a hands-on KingBench run put it sixth of eleven. Independent timing shows it is slow, and the 2-3x Highspeed claim has not been timed by anyone we found. Community reports are mixed: instruction following with a system prompt drew complaints, and subscribers say usage runs out quickly. Rough edges are written down here; the promise that it is a frontier-class open-weight model that ships when said holds. Re-stamped on day 15 (7 Nov).

Launch dayWhat they claimed, quoted word for word · 4

“Kimi K3.1 Pro and Kimi K3.1 Pro Highspeed are available today on Kimi.ai, Kimi Work, Kimi Code, and the Kimi API. The full model weights will be released by Nov 4.”
✓ HoldsEvidence: R3, R5, S1, S2

Hosted access was live on launch day and the weights were published on Hugging Face on 4 Nov, inside the promised date. The release ships no modelling code, which one porting team flagged.

“our most capable 3T-class open-weight model to date, with a 1-million-token context window”
✓ HoldsEvidence: S1, S2

Artificial Analysis lists 2.8T total parameters and a 1M-token context window, and the model card gives the same context length.

“K3.1 Pro scores 65.7%, ahead of GPT-6 Astra (58.2%) and narrowly behind Claude Opus 5.5 (66.4%)”
~ MixedEvidence: R6, S1, S2, S3, S4

Moonshot ran its own model in the Claude Code harness and took rivals' numbers from published reports. Artificial Analysis has since moved the model to 46 on its revised index, and a hands-on benchmark ranked it sixth of eleven. A community arena has it tied with DeepSeek V4.1 Flash.

“Kimi K3.1 Pro Highspeed, which delivers 2-3× the speed of K3.1 Pro for more responsive interactions”
? UnclearEvidence: R1, R2, R7, S2, S3

Artificial Analysis measures the standard Pro endpoint at 33 tokens per second and calls it notably slow per task. We found no independent timing of Highspeed.

Day 7What real users report · 7

R1Hacker News24 Oct 2026mixed

A tester who paid for the higher plan says the output quality is solid but the five-hour usage limit still runs out too fast, and ranks the model fourth behind Opus 5.5, Astra and Argon.

Open the source

R2Hacker News23 Oct 2026praise

A user says the model is slower than rivals but follows rules better and makes fewer silly mistakes, so solutions need less time to merge.

Open the source

R3Forum5 Nov 2026praise

A team ported the released weights to its own inference engine in a day: tensor names match the previous K3 apart from a new indexer, and the 4-bit experts repack losslessly. They note the repository has configuration but no modelling code, so part of the indexer had to be reconstructed.

Open as of 7 Nov · 1 post · Open the source

R4Forum4 Nov 2026problem

A user building a research assistant says the model keeps falling back to its built-in Kimi persona and ignores their system prompt, which other open models respect. Other commenters ask for more flexibility.

Open as of 7 Nov · 11 comments · Open the source

R5Forum6 Nov 2026praise

Two days after release someone shares a project that runs the full model locally on a single 96 GB GPU, at about 2 tokens per second, with prefill at about 40 tokens per second.

Open as of 7 Nov · 2 comments · Open the source

R6Hacker News4 Nov 2026mixed

The builder of an agent arena says K3.1 and DeepSeek V4.1 Flash are basically tied on their leaderboard, with K3.1 slower and more expensive, though the error bars are big.

Open the source

R7Hacker News4 Nov 2026problem

A commenter says the subscription lasts only minutes at a time, API costs are very high, and the model is neither faster nor cheaper than an Anthropic subscription, unlike the Flash-class alternatives.

Open the source

OthersWhat outlets and testers say · 4

S1Simon Willison23 Oct 2026positive

Tried it through the API on launch day, calls it the first soon-to-be open-weight model to score this close to the frontier, and notes the pelican test is solid apart from both legs on one side.

“Introducing Kimi K3.1 Pro” · Open the source

S2Artificial Analysis7 Nov 2026mixed

Open weights, 1M context, 46 on its current Intelligence Index, but notably slow per task (36.1 minutes), slow to the first answer token, and pricey versus other open-weight models.

“Kimi K3.1 Pro: Intelligence, Performance & Price Analysis” · Open the source

S3Leiphone (via NetEase)2 Nov 2026mixed

Ran 12 agent tasks in a week: capabilities broadly accepted, but the tasks ran slowly, the third-party KingBench ranks it sixth of eleven, and Kimi apologised for throttling and odd quota burn in the first days.

“Kimi K3.1 Pro hands-on: top marks in science, a step behind in the humanities” · Open the source

S4Agent Native Engineering29 Oct 2026mixed

Explains that the independent score fell from 55 to 46 when the index was reweighted toward private tests, that Moonshot's table mixes harnesses, and that cost per task stays well below Opus 5.5.

“Kimi K3.1 Pro scored 55, then 46 four days later” · Open the source


Claims ✓2 ~1 ✗0 ?17 reports
First stamped 7 Nov 2026Last re-stamped 7 Nov 2026