Every figure quoted in the video, where it came from, and how it was checked. Where sources disagree, the disagreement is written down rather than resolved silently.
Checked on 30 July 2026.
| Claim | Figure | Source |
|---|---|---|
| Developer | Moonshot AI, Beijing | Model card |
| API available | 16 July 2026 | VentureBeat |
| Open weights published | 26 July 2026 | roo.beehiiv.com |
| Date Moonshot had promised | 27 July 2026 | Same |
The ten day gap between the API opening and the weights being published is the difference between those first two dates. The weights landed roughly one day ahead of the 27 July date Moonshot had given.
Kimi K3 is the largest open weight model published to date at 2.8 trillion total parameters. The previous largest was DeepSeek V4 Pro at 1.6 trillion, which makes the increase about 1.2 trillion parameters.
All from the model card and the Kimi K3 tech blog.
| Property | Value |
|---|---|
| Total parameters | 2.8T |
| Activated parameters per token | 104B |
| Experts | 896 total, 16 active per token |
| Layers | 93 (69 Kimi Delta Attention, 24 Gated MLA) |
| Context length | 1,048,576 tokens |
| Vocabulary | 160K |
| Vision encoder | MoonViT-V2 |
| Modalities | text, images and video in one model |
| Weight format | MXFP4 weights, MXFP8 activations |
| Optimizer | Per Head Muon |
Moonshot’s tech blog claims roughly a 2.5 times improvement in overall scaling efficiency compared with Kimi K2.
One caveat worth stating: some coverage described the active compute as “roughly 50 billion parameters” by multiplying 16 of 896 experts against the total. The 104B figure used in the video is the one Moonshot publishes on the model card, and it is the higher of the two because attention and shared parameters are active on every token regardless of which experts are routed to.
This is the number most coverage got wrong, so it is worth setting out in full.
What the repository actually contains. The Hugging Face repository lists a total size of 1.56 TB across 96 safetensors shards. Repository file listing
Why 594 GB cannot be the model. 2.8 trillion parameters at four bits each is 1.4 terabytes before anything else is added. 594 GB would work out at under two bits per parameter, which does not match a model shipped in MXFP4.
Where 594 GB came from. It is the size of a community produced one bit dynamic quantization of the model, published by Unsloth, measured between 594 GB and 620 GB across same day revisions. The figure was picked up as though it were the release itself. Run Kimi K3 locally, hardware reality check
The rest of the ladder, from the same source:
| Build | Size |
|---|---|
| One bit dynamic | 594 GB to 620 GB |
| Two bit dynamic | 711 GB |
| Four bit | 1.51 TB |
| Eight bit | 1.56 TB |
At full precision. At the 16 bit precision most models ship in, 2.8 trillion parameters would come to roughly 5.6 terabytes. Moonshot never shipped that: the model was trained with quantization aware training from the supervised fine tuning stage onward, so the four bit version is the model rather than a compressed copy of one. Kimi K3 tech blog
A note on 1.4 TB versus 1.56 TB. Both figures appear in coverage and both are right about different things. 1.4 TB is the weights themselves at four bits. 1.56 TB is the total size of the repository as Hugging Face reports it, which includes everything else in it. The video uses 1.56 TB when talking about what you download and 1.4 TB when talking about what has to fit in memory.
| Claim | Figure | Source |
|---|---|---|
| Accelerators needed to hold the weights | around eighteen 80 GB cards | TECHi |
| Newest single node | 8 cards at 192 GB, about 1.5 TB total | Same |
| Production serving | 64 or more accelerators | Same |
| Memory floor for the one bit build | about 650 GB of combined RAM and VRAM | modemguides |
| Reported consumer path | a 512 GB Mac Studio paired with a second machine | Same |
| One bit build on data center cards | more than 100 tokens per second on B200 class hardware | Same |
TECHi’s framing of the binding constraint, that it is memory capacity and bandwidth rather than floating point throughput, and that inference cost is increasingly measured in tokens per megawatt, is quoted in the video as their analysis rather than as a measurement.
This is the second claim most coverage got wrong.
Kimi K2 shipped under a Modified MIT licence whose only additional condition was attribution at very large scale.
Kimi K3 does not. The file in the repository is a bespoke document titled the Kimi K3 License, not a variant of MIT, and it adds a condition K2 never had. LICENSE
The two thresholds, as written:
The licence explicitly carves out internal use that is not accessible to third parties, and access through Moonshot AI’s own products or its certified inference partners. That is why the video describes the first threshold as aimed at resellers rather than at users: modifying, distributing, building products on and running the model inside a company all remain permitted without a separate agreement.
The distinction between “open weight” and “open source” matters here, and the video does not call the model closed. It is freely downloadable and freely usable by the overwhelming majority of people who would want it. It is not unconditionally free for a business whose product is reselling it.
Comparison of the K2 and K3 terms
Independent measurements from Artificial Analysis:
| Measure | Kimi K3 |
|---|---|
| Intelligence Index | 57 |
| Rank among open weights models | 1st of 98 |
| Output speed | 34.4 tokens per second |
| Time to first token | 3.36 seconds |
Artificial Analysis describe K3’s intelligence as comparable to Claude Opus 4.8 and GPT 5.5, and behind Claude Fable 5 and GPT 5.6 Sol. The two published index scores above K3 are Fable 5 at 60 and GPT 5.6 Sol at 59, which is where the three point gap quoted in the video comes from. Artificial Analysis writeup
Because Artificial Analysis do not publish separate index scores for Opus 4.8 and GPT 5.5 alongside K3’s, the video shows those two as a comparable band around 57 rather than assigning them numbers.
Scores from Moonshot’s own model card:
| Benchmark | Score |
|---|---|
| GPQA Diamond | 93.5 |
| Terminal-Bench 2.1 | 88.3 |
| BrowseComp | 91.2 |
| DeepSWE | 67.5 |
| OSWorld 2.0 | 58.3 |
Agentic work. Artificial Analysis report a rating of 1668 on their professional task evaluation, up from 1190 for K2.6. One other summary of the same evaluation quoted 1687; the video uses the lower and more widely reported 1668, and the point being made, that this is a large single generation jump, holds either way.
Front end code arena. K3 was ranked first at 1679, ahead of Claude Fable 5. Tom’s Hardware. Only the leading score is published, which is why the video shows the models below it in position without scores. Arena rankings measure human preference on a particular kind of task and move around; the video flags this rather than treating the ranking as settled.
From Artificial Analysis:
| Measure | Kimi K2.6 | Kimi K3 |
|---|---|---|
| Hallucination rate | 39% | 51% |
| Accuracy rate | 33% | 46% |
Both moved in the same direction, and that is the point. The model answers more questions correctly than its predecessor, and it also declines to answer less often, which produces more confident wrong answers alongside the extra right ones. These are two different measures on the same evaluation, not two halves of one number.
| Model | Input, cache miss | Input, cache hit | Output |
|---|---|---|---|
| Kimi K2.6 | $0.95 | $0.16 | $4.00 |
| Kimi K3 | $3.00 | $0.30 | $15.00 |
| Claude Sonnet 5 | $3.00 | $15.00 | |
| Claude Fable 5 | $10.00 | $50.00 |
All per million tokens. the-decoder, cross checked against Artificial Analysis.
K3’s input price is roughly three times K2.6’s, and lands at exactly the same headline figures as Claude Sonnet 5. The cache hit price of 30 cents remains substantially cheaper than the western equivalents. The wider argument, that Chinese frontier models are no longer priced at a steep discount to western ones, is the-decoder’s and is presented as such.