NV-Reason-CT vs Kimi K3: 210 Downloads and 588 Million Tokens

One of these models has been downloaded 210 times from Hugging Face. The other moved 588 million tokens through our own playground in the last seven days. Both are open-weights, both are downloadable right now, and the gap between those two numbers is the most useful thing in this comparison. NV-Reason-CT is NVIDIA's 4.69B-parameter 3D vision-language model for chest and abdominal CT, published to Hugging Face on September 8, 2026 with no announcement. Kimi K3 is MoonshotAI's 1,048,576-token flagship, shipped July 15, 2026 and now one of the busiest models on our network. They are both "open" in the sense that matters to a procurement form. They are not open in the same way at all.

Two meanings of "you can download it"

Start with what each download actually deposits on your disk, because this is where most open-weights comparisons go soft.

NV-Reason-CT gives you model.safetensors at roughly 10.6 GB in BF16, plus a model.py and a substantial processor.py — 26 KB of custom processing code that implements the anatomy-aware cropping, the 2-mm resampling and the 3D patchification. You load it with trust_remote_code=True, which means you are executing NVIDIA's repository code inside your process. Inference needs CUDA PyTorch, and the card lists Ampere, Hopper and Lovelace, tested on H100 and L40S. Input is one NIfTI volume at a time. There is no batching story published, no throughput figure, and no serving configuration in the repository beyond an inference.py example.

Kimi K3 gives you a general-purpose reasoning model with a 1,048,576-token context window, text and image input, native tool calling, structured outputs, and configurable reasoning effort. It is the model you would reach for to read a hundred clinical guidelines in one request, or a decade of a department's reports. Its Artificial Analysis long-context recall score is 88.7 — the highest of the models in this series and a full 8.4 points above Grok 4.6's 80.3.

Put plainly: NV-Reason-CT is a specialist checkpoint with a narrow input surface and custom code you must trust and maintain. Kimi K3 is a general model with a wide input surface that behaves like infrastructure. Both are downloadable. Only one of them is a drop-in.

The CT evidence, in full, and it is short

Everything publicly known about NV-Reason-CT's quality rests on a single table in its own model card: CT-RATE's 18-label classification task at a fixed uniform threshold, direct Yes/No prompting, no classification head, no task-specific adaptation. Macro-F1 0.614, Macro-AUROC 0.871.

The comparison rows are VoxelFM at 0.581/0.870, Pillar-0 at 0.544/0.861, ClinFusion-8B at 0.442, CT-CLIP at 0.398/0.733, Merlin at 0.358/0.662 and MedGemma 1.5 at 0.303. Every one of those is a vendor-reported number from NVIDIA's card and none has been independently reproduced yet — the paper is two days old.

What the table establishes is narrow and worth stating exactly: a generative 3D model with no task-specific head beats 3D contrastive pretraining baselines and a slice-based frontier model at 18-label CT classification. What it does not establish is anything about general reasoning, anything about modalities other than chest and abdominal CT, and anything about the AUROC advantage — 0.871 against 0.870 is a tie wearing a decimal point.

Kimi K3's numbers, and the open-weights board caveat

Artificial Analysis measures Kimi K3 at an Intelligence Index of 43.6, which places it 19th of 145 models on that board in our model API's snapshot of the ranking. Its GPQA Diamond score is 93.5, Humanity's Last Exam 46.9, SciCode 59.5, Terminal-Bench 2.1 85.0, AA Coding 76.2 at rank 9, and 𝜏²-banking 46.0.

One ranking claim needs care, because it is easy to get wrong and it has been got wrong before. Kimi K3's AA Intelligence Index of 44 sits third among open-weights models on the open-weights board, behind MiMo-V2.6-Pro at 46 and GLM-5.3 at 45. It is not the leader on that board, and it is not competing on the same board as Grok 4.6 — Artificial Analysis records Grok 4.6 as proprietary, so its 44 belongs to the general ranking, not the open-weights one. The two indexes look identical and are not.

What the operator actually sees

This is where the 588 million number comes from, and it is worth spelling out because it is the one piece of evidence in this article that is ours rather than anyone's marketing.

Over the last seven days, Kimi K3 moved 587.8 million tokens through OrcaRouter's playground. Its median time to first token was 7.8 seconds, it produced 42.9 output tokens per second, and its error rate over that window was 0.21% — one failure in roughly 480 requests. That measured error rate is the thing you cannot get from a model card, and it is the number that decides whether a model is safe to put behind a batch job.

For NV-Reason-CT there is no equivalent, because it is not a hosted endpoint anywhere — it is a checkpoint you run yourself. There is no p50, no error rate, no uptime, and no one to failover to. That is not a knock; it is a structural fact about self-hosting, and it is the reason a research group can adopt NV-Reason-CT on a Tuesday and a production team generally cannot.

If you are running both halves of this — Kimi K3 as the hosted general layer and NV-Reason-CT on your own GPU box — the routing side is where the hosted half gets its resilience. Kimi K3 is served at Moonshot's own list price of $3.00 per million input and $15.00 per million output with zero markup, its cache-read rate of $0.30 per million is a tenth of input, and because the price is passed through unchanged a vendor cut reaches your bill the same day. Automatic failover across providers is what keeps a batch job alive when one upstream endpoint wobbles — and at a 0.21% error rate, wobbles are rare enough that they are easy to forget about until they are not.

The cost shapes do not resemble each other

• Price — NV-Reason-CT GPU capex plus your own serving stack vs Kimi K3 $3.00 / $15.00 per million, $0.30 cached read

• Minimum viable spend — NV-Reason-CT one Ampere-or-newer GPU held indefinitely vs Kimi K3 one request

• Context — NV-Reason-CT one volume, 13,824 tokens fixed vs Kimi K3 1,048,576 tokens

• Marginal cost of the thousandth request — NV-Reason-CT effectively zero once the GPU is paid for vs Kimi K3 linear in tokens

• Where it breaks first — NV-Reason-CT unverified throughput and no batching story vs Kimi K3 long-context cost at 1M tokens per request

• Modalities — NV-Reason-CT 3D NIfTI, chest or abdomen vs Kimi K3 text and image, anything

The economics invert depending on volume. Kimi K3 costs nothing to start and scales linearly. NV-Reason-CT costs a GPU to start and then scales almost flat — which is why 210 downloads is not a failure signal. It is the correct number for a checkpoint whose audience is institutions that already own the hardware, and who will never appear in a token-throughput statistic at all.

Which one you are actually choosing

If the job is reading a CT volume and producing a structured finding with a reasoning trace, Kimi K3 cannot do it and NV-Reason-CT can. Not "does it worse" — cannot, because there is no volumetric encoder anywhere in it, and flattening a scan into images first discards the spatial relationships the 3D pathway exists to preserve.

If the job is reading 400 pages of referral notes, matching them against a protocol document, and emitting structured output, NV-Reason-CT is not in the conversation and Kimi K3 is one of the better answers available, at a measured error rate under a quarter of a percent on our network.

The genuine decision is not between the two. It is whether your problem is a volume problem or a text problem, and the 210-versus-588-million gap is simply what that distinction looks like from the outside. One model is a paper with weights. The other is a piece of infrastructure. Both are worth having; neither substitutes for the other.