Someone got Kimi K3 running on a MacBook Pro. 2.78 trillion parameters, MXFP4 experts streamed off four SSDs, at 1.00 token per second. On a 512-token prompt, time to first token was about 6.3 minutes.
I read that thread twice. Not for information. Because it is funny in a way that only hardware people will find funny. The model is smarter than anything i will ever run at home, and running it at home is technically possible, the way driving to another city in first gear is technically possible.
Then Moonshot shipped Kimi K3.1 Pro on October 23 and the argument changed shape.
Here is the thing that caught me. K3.1 Pro jumped from 43.6 to 55.1 on the Artificial Analysis Intelligence Index. Same base weights underneath. Fourth of 215 models, above GPT-6 Astra, below Claude Sonnet 5.5 and Opus 5.5.
A lab took a model it had already released, spent more on post-training, and moved it eleven points. That is not supposed to be the easy direction this late.
But the loudest complaint on release week was not about capability. It was about quota.
What actually changed
K3.1 Pro is not a new model. It is Kimi K3 with a much larger post-training budget bolted on.
The architecture is identical. 2.8 trillion parameters, 896 experts, 16 active per token, one-million-token context. Kimi Delta Attention, Attention Residuals, all of it carried over untouched.
So when someone tells you the base model got smarter, that is not what happened.
The base model got a better education.
What they actually changed was the RL stack. Moonshot says it rebuilt:
- the RL framework
- long-context inference
- sandbox infrastructure
- rollout scheduling
- a shared platform for checking environment quality
That is a boring list and it is the whole release. They built the factory, then ran it longer.
the interesting claim is not that the model is good. it is that more post-training compute now returns stable gains instead of noise.
If that holds, every lab with a big base model and a training budget has a cheaper path to the top of the board than they thought.
The numbers are real, but lopsided
The coding jump is the one Moonshot leads with. On Terminal-Bench 4.0 it scores 65.7%. GPT-6 Astra is at 58.2%, Claude Opus 5.5 is at 66.4%.
On SWE Marathon, a multi-hour engineering run, it goes 42.0 to 51.7. On the internal Kimi Code Bench long-horizon test, 55.95% to 61.2%.
And the hallucination number is the one i did not expect. On Kimi Hallu Bench Hard, built to hit edge cases rather than normal use, the error rate dropped from 23.4% to 8.3%.
That matters more than the benchmark wins. A model that stops inventing things is worth more than a model that scores two points higher.
Now the other half of the table.
| Benchmark | K3.1 Pro | Opus 5.5 |
|---|---|---|
| Terminal-Bench 4.0 | 65.7% | 66.4% |
| BenchCAD | 92.7% | 96.2% |
| TB Science | 38.1 | 25+ pts ahead |
Terminal-Bench Science at 38.1 is a cliff, not a gap. HLE moved 47% to 52%, CritPt 23% to 25%, SciCode 59% to 61%.
The split lines up with something obvious. Agents and code are where RL hands you cheap, verifiable reward. Science is where you cannot just check the answer.
The token bill nobody priced in
K3.1 Pro runs at about 49.7 tokens per second on Artificial Analysis. That ranks it 165th of 182.
That alone would not kill it. Patient background jobs do not care about latency.
The real problem is volume. Running the AA index took 150 million output tokens against a peer median of 87 million. Roughly 72% more talking than the average model on the same tasks.
Combine those and the price story gets weird. Input is $2 per million, output is $10, a third cheaper than K3. Sounds great.
But if the model spends 72% more tokens to finish, you gave a lot of that discount back.
Cheap per token is not the same as cheap per task. Measure it on your own workload before you move anything to kimi-k3.1-pro.
Then there are the subscribers. Multiple people on Kimi Code plans reported burning a five-hour limit in under an hour. Whether that is retry storms, verbose output, or something in the plan math, i do not know. But it was the complaint that showed up most in week one.
The part Moonshot buried
Here is the detail almost nobody wrote about.
Before shipping the open weights, Moonshot removed offensive cyberattack training data from both stages of post-training. Not an output filter. A change to the training data itself.
They measured it. K3.1 Pro scores 81.0% on their in-house xdev-bench-hard cyber evaluation. The same model trained without the redaction scores 90.5%.
They published the 9.5-point capability cut on purpose and said it keeps K3.1 Pro below the offensive cyber level of already-public proprietary models.
I have not seen another open-weight lab document a voluntary capability limit like that. Especially not one shipping to a global audience while the US government accuses it of distilling American models.
I am not naive about this. A published safety ablation is also excellent press.
But it is checkable. That is more than most safety claims get.
Speed is the tax
The Highspeed variant is the tell. Moonshot shipped a companion model that runs 2 to 3 times faster, priced at $4 input and $20 output, because the base model is slow enough that people asked for one.
You do not build a fast lane unless the normal lane has a traffic problem.
None of this is why i have four external SSDs on my desk.
I bought the first one for a project i finished. The second for a backup of the first. The third because i read a thread about running big models off disk and thought, that could be me. The fourth i genuinely cannot explain.
I am not running a 2.8 trillion parameter model off them. I am running the same two folders i have been shuffling between drives since 2023.
There is a specific kind of person who reads "1 token per second from four SSDs" and thinks, that is my weekend. I am that person. I have never once actually done it.
The drives stay plugged in. They hum. I keep reading about quantization.
Most people do not need K3.1 Pro.
If your work is chat, autocomplete, or anything a human sits and waits on, 49 tokens per second will annoy you and the extra tokens will cost you. Use something fast.
If your work is a support agent answering tickets, the model is the smaller half of your problem anyway. Retrieval and setup decide more than the leaderboard does.
And if you planned to self-host next week, the weights are not out. Moonshot says by November 4. That is a promise, not a file.
The honest case is narrow. Long unattended agent runs, where you leave it working and check back later. Near-frontier quality at a third off K3's price actually pays there.
Everything else, wait for independent reruns. The vendor numbers are good. They are still vendor numbers.
What i am actually watching
I keep coming back to that MacBook thread.
Someone spent a weekend getting 1 token per second out of a model that would have been a research paper three years ago. Not useful. Not scalable. But it worked, and they wrote it down.
A week later the same lab shipped a version that ranks fourth in the world, and the top comment was about quota limits.
Both of those are the story. The model got better. The packaging did not keep up.
Whichever one Moonshot fixes first tells you what they actually care about.
Top comments (0)