Kimi's 1.56 TB download lost 20 rows to Opus

Moonshot's Kimi K3.1 Pro weighs 1,561 GB across 96 files, claims 2.8 trillion parameters — and its own README table shows it losing to Claude Opus 5.5 on 20 of the 39 rows where both have scores. The download is real. The "best open model" framing does not survive the vendor's own numbers.

I did not download 1.56 TB or run the model. What I did is audit everything that can be checked without that: the Hugging Face file listing, the config, the license, and all 46 rows of the evaluation table the model card ships. Every number below is from those sources or arithmetic on them.

The release, measured

The repo (moonshotai/Kimi-K3.1-Pro) was created November 2, 2026. Three days in it has 11,658 downloads and 801 likes. The file listing tells the story before any benchmark does:

safetensors shards:   96
total size:           1,561.29 GB
median shard:         16.99 GB
largest shard:        16.99 GB
smallest shard:       0.09 GB
model.safetensors.index.json: 59.77 MB
generation_config.json:       53 bytes

Ninety-five of the shards are ~17 GB each; the last is a 0.09 GB tail. The 60 MB index file is the map for 2.8 trillion parameters. The 53-byte generation config is almost empty — there is barely anything to configure at load time.

The config itself:

total parameters:     2.8T (per README)
activated per token:  104B (per README)
layers:               93 (1 dense, 92 MoE)
experts:              896, 16 selected per token
attention:            69 KDA + 24 Gated DSA layers
hidden size:          7168
context length:       1,048,576 tokens
quantization:         MXFP4 weights / MXFP8 activations, QAT
vision encoder:       MoonViT-V2, 401M params

The byte arithmetic checks out. At 4 bits per weight, 2.8 trillion parameters need 1.4 TB. The download is 1.56 TB, and the gap is explained by the config's own quantization ignore-list: attention layers, shared experts, the lm_head, and the vision tower stay unquantized, plus per-group scales. That is the honest reading — the size is consistent with the claimed parameter count, not inflated by it.

One more check on the "3T-class" label in the README: 2.8T rounds to 3T only with generosity. It is a 2.8T model that markets itself in the next class up. Keep that in mind for the table below.

The evaluation table, counted

The README ships a 46-row evaluation table against five other models: Claude Fable 5.1, GPT-6 Astra, Claude Opus 5.5, DeepSeek-V4.1-Flash, and GLM-5.3. All Kimi scores use reasoning effort max, temperature 1.0, top-p 0.95. I parsed every row and counted wins and losses instead of reading the headline.

On the 44 higher-is-better rows:

vs Claude Opus 5.5:  19 wins, 20 losses (5 rows have no Opus score)
vs GPT-6 Astra:      22 wins, 16 losses (6 rows have no Astra score)
best of all 6 models: 16 of 44 rows

Kimi loses to Opus more often than it beats it. That alone should end any "SOTA" claim, but the individual gaps are the real story:

Program Bench:            Kimi 31.5  vs Opus 65.0   (less than half)
HLE-Diamond:              Kimi 39.4  vs Astra 65.8
Terminal-Bench-Science:   Kimi 38.1  vs Astra 68.1
ExploitBench:             Kimi 74.7  vs Astra 100.0
HLE-Full:                 Kimi 51.1  vs Opus 59.8
Terminal-Bench 4.0:       Kimi 65.7  vs Opus 66.4   (closest headline loss)
DeepSWE v1.1:             Kimi 75.5  vs DeepSeek 77.0

The wins are real too — SWE Marathon v1.1 (51.7, best of six), BrowseComp (92.1, best), HLE-Full with tools (70.0, best), Terminal-Bench 4.0 is a near-tie. The pattern is not "Kimi is bad." It is "Kimi is a strong open model that wins the agent-with-tools rows and loses the hard reasoning rows" — which is the opposite of the launch framing. The two lower-is-better hallucination rows go Kimi's way (8.3 and 5.4, both best), so the hallucination story is the strongest part of the table.

Two more table details worth stating plainly. Eight of the 46 rows are marked Internal — Moonshot's own benchmarks, where the methodology is theirs alone. And two rows have dashes for all three closed models, so there is nothing to compare against at all: Kimi Code Bench 2.0 (77.1) and xdev-bench-hard (81.0) are Kimi-only numbers. A table where your best-looking rows are the ones nobody else was scored on is a table doing marketing work.

The license is not open source

The LICENSE file is a custom "Kimi K3.1-Pro License," not MIT or Apache. It reads like MIT for the first paragraph — use, copy, modify, distribute, sublicense, sell — then adds two gates:

1. If you run it as a Model-as-a-Service business and your
   aggregate revenue exceeds $20M over any 12 months, you must
   sign a separate agreement with Moonshot AI before ANY
   commercial use.
2. If your product has >100M monthly active users or >$20M in
   monthly revenue, "Kimi K3.1-Pro" must be displayed
   prominently in the UI.

Internal use is exempt. So is use through Moonshot's own products and certified inference partners. For everyone else, the weights are free until you are big enough to matter, and then they are not. That is a reasonable business decision and a real restriction — "open weights" is the accurate label, not "open source."

The verdict

Three days after release, the facts that check out are: 96 shards, 1,561 GB, 2.8T parameters at MXFP4, 1M-token context, 11,658 downloads. The facts that do not check out are the framing: on its own table, Kimi K3.1 Pro beats Opus 5.5 on 19 rows and loses on 20, wins 16 of 44 outright, and its worst gaps are on the hardest reasoning benchmarks, not the easiest ones. If you want a 2.8T open model for agentic coding with tools, the download is justified. If you want the model that beats the frontier on reasoning, the vendor's own table says to look elsewhere — and it says so in the rows most people will not read.

Sources: Hugging Face repo moonshotai/Kimi-K3.1-Pro (file listing via API, November 5, 2026; config.json; README.md evaluation table; LICENSE). All counts above are my own parse of the README table; shard sizes are from the Hugging Face tree API. I did not download the weights or run the model.