The first line of the NV-Reason-CT model card tells you something most people skim past. The base model is Qwen/Qwen3.5-4B, listed in the repository metadata as a fine-tune relation. So when NVIDIA needed a language model to bolt a Primus 3D vision transformer onto, it took a Qwen. Compare that checkpoint with Qwen3.8 Max — Alibaba's own 1,000,000-token flagship, shipped August 3, 2026 — and you are looking at a fine-tune and the family business. One is a 4.69B specialist that reads chest and abdominal CT volumes and was published to Hugging Face on September 8, 2026 with no announcement. The other is a general flagship with text, image and video input, a published rate card, and 37.2 million tokens of measured traffic on our network in the last week. The interesting question is not which wins. It is what a 4B fine-tune can do that its family's flagship cannot, and why the answer is more specific than you would expect.
What the fine-tune bought, concretely
NVIDIA's own framing of the gap is unusually precise, and it is the most useful sentence in the repository: most vision-language systems "either operate on 2D images or compress volumetric features before language decoding." That is a description of an entire class of models, and it includes every general multimodal flagship on the market.
The fix is architectural rather than a matter of scale. A 384×384×384-mm input volume becomes a 24×24×24 grid of 13,824 visual tokens. Those tokens and their three-dimensional coordinates go into the language model with no spatial downsampling, and 3D MRoPE preserves the spatial relationships inside the decoder. Input volumes are cropped anatomy-aware to chest or abdomen using lung Hounsfield units, then resampled to 2-mm isotropic resolution.
Read that against what Qwen3.8 Max accepts. Text, image, and video — and video is the closest thing in this series to a volumetric input, which makes it the honest comparison point rather than a rhetorical one. A video is a stack of 2D frames ordered by time. A CT volume is a stack of 2D slices ordered by physical position along the z-axis, in Hounsfield units, with a known millimetre spacing. Both are three-dimensional arrays. Only one of them has a physical coordinate system that a radiologist's finding can be localised to, and only one of them is measured in a unit that means anything outside an image sensor. A model trained on video learns temporal continuity. A model trained on CT volumes learns that a mass in one slice is the same mass in the next slice eight millimetres lower.
The training corpus is the real difference
This is where the two models stop being comparable at all, and it is the part a spec sheet hides.
NV-Reason-CT was trained end-to-end with supervised fine-tuning followed by Group Relative Policy Optimization on roughly 550,000 structured QA examples drawn from 70,111 unique CT volumes. The sources are CT-RATE — 47,149 cases from Istanbul Medipol University Mega Hospital with paired radiology reports — an internal NIH set of 15,991 cases, and CancerVerse's 22,720 cases across four contrast phases and thirteen malignant tumour types. The corpus includes standardized reports, abnormality-focused QA, multi-turn dialogue, and radiologist-authored reasoning transcribed from recorded expert interpretations. The GRPO stage uses verifiable rewards over chest and abdominal abnormality sets, which means the reward signal checks whether a stated finding is present in the volume rather than whether the prose sounds clinical.
Qwen3.8 Max's training corpus is not published in anything like that detail, and it would not help if it were, because the target capability is general. Its Artificial Analysis figures are the relevant evidence instead: an Intelligence Index of 45.4 placing it 15th of 145 models in our API's snapshot, AA Coding 76.2 at rank 9, Terminal-Bench 2.1 at 88.8, GPQA Diamond 92.8, Humanity's Last Exam 43.1, SciCode 52.1 and long-context recall 80.3. Those are third-party measurements, not vendor claims, which makes them the most trustworthy numbers on either side of this page.
Reasoning output, two different contracts
Both models emit reasoning traces and both let you turn them off. The similarity ends at the interface.
Qwen3.8 Max exposes enable_thinking and reasoning_effort, along with the full sampling surface — temperature, top_p, seed, penalties, structured outputs, tool calling. It is a general reasoning capability you point at whatever you have. Its thinking is as good as the problem you hand it, and it has no way to verify a claim against anything except the text it was given.
NV-Reason-CT's reasoning has a narrower contract with the reader, and the model card states the limit explicitly: generated reasoning is "reviewable model output and is not guaranteed to represent the model's internal computation." That caveat is doing real work, and it is the sort of sentence that only appears when a team has thought about what a clinician will do with a plausible-sounding trace. The card describes the output as reviewable observations, differential diagnoses and uncertainty. A trace you can check against the volume, in other words, rather than one you can only read.
Throughput, from the one place we can measure it
Qwen3.8 Max moved 37.2 million tokens through OrcaRouter's own playground in the last seven days. Its median time to first token was 1.96 seconds — comfortably the fastest of the models in this series — at 53.3 output tokens per second. Its error rate over that window was 3.5%.
Those two figures pull in opposite directions and both matter. The latency is excellent and makes the model pleasant to iterate against. The error rate is roughly seventeen times higher than Kimi K3's 0.21% over the same window, and that is the number that decides whether you put it behind a synchronous user-facing call or a batch job with retries. This is measured behaviour on our network, not a benchmark, and it is the kind of thing a rate card never tells you.
NV-Reason-CT has no equivalent measurement anywhere. It is a 10.6 GB BF16 checkpoint with custom model code, loaded with trust_remote_code=True on Ampere, Hopper or Lovelace hardware, tested by NVIDIA on H100 and L40S. There is no p50, no error rate and no throughput figure published. Anyone who tells you the crossover volume at which self-hosting this beats paying per token is guessing.
Price and shape, side by side
• Price — NV-Reason-CT GPU capex, no per-token rate vs Qwen3.8 Max $2.00 / $6.00 per million tokens, $0.25 cached read, $2.50 cache write
• Input — NV-Reason-CT 3D NIfTI CT volumes in Hounsfield units, chest or abdomen only vs Qwen3.8 Max text, image and video
• Context — NV-Reason-CT one volume, a fixed 13,824-token prefix vs Qwen3.8 Max 1,000,000 tokens
• Parameters — NV-Reason-CT 4.69B total / 4.35B active in the CT pathway vs Qwen3.8 Max undisclosed
• Licence — NV-Reason-CT OpenMDW-1.1, weights published vs Qwen3.8 Max proprietary, API access only
• Independent scores — NV-Reason-CT none vs Qwen3.8 Max AA Intelligence Index 45.4, GPQA Diamond 92.8
Qwen3.8 Max's cache economics reward a specific shape: a $0.25 cached read against $2.00 fresh input is an eightfold discount, and the $2.50 cache write means a long reusable prefix pays for itself quickly. That is the workload of a system prompt plus a protocol document reused across thousands of requests. NV-Reason-CT has the opposite cost profile — near-zero marginal cost per scan once the GPU is bought, and a hard floor of one GPU held indefinitely.
Running the family together
The natural architecture here, and it is worth saying that nobody has published it, is the fine-tune for the volume and the flagship for everything around it. NV-Reason-CT produces the structured finding; Qwen3.8 Max handles the referral text, the protocol matching, the coding, the follow-up letter — the high-volume text work where a million-token window and an eightfold cache discount actually earn their place.
If that is your pipeline, half of it is a checkpoint you host and half is tokens you buy, and the second half is where a router does something a single vendor endpoint cannot. Qwen3.8 Max is served through OrcaRouter at Alibaba's list price with zero markup, so the $2.00 / $6.00 above is what you pay and any future price move lands on your bill the same day rather than at renewal. Automatic failover across providers matters more than usual when the model's own measured error rate is 3.5% — a retry that goes to a different healthy endpoint is the difference between a transient blip and a failed batch. And because the general half of the pipeline is one model name behind an OpenAI-compatible endpoint covering 200+ models, swapping in something else for a week's experiment costs a string change, not an integration.
NV-Reason-CT, to be clear, is not on OrcaRouter and is not going to be — it is custom-code 3D inference you run yourself. The honest version of the integration story is that the self-hosted specialist and the routed generalist can sit under one key, and that is the whole of the claim.
The verdict that actually holds
This pairing gets filed under "which is better" and it should not. Qwen3.8 Max is measured by third parties on general reasoning and beats most of the field; NV-Reason-CT has one table of its own numbers on one CT benchmark, and a 4.69B parameter count that would be unremarkable in any general comparison. On any general benchmark, the flagship wins, and the reason is that the general benchmark is what it was built for.
On the one task NV-Reason-CT was built for — reading a chest or abdominal CT volume and saying what is in it, with a trace someone can check — Qwen3.8 Max is not a weaker competitor. It has no volumetric encoder, so it cannot be run on the input at all without someone first deciding how to flatten a scan into pictures. That decision is where the spatial information goes. That is the entire point of a 4B fine-tune with a native 3D pathway and a 13,824-token fixed prefix, and it is why the interesting thing about this model is its backbone's smallness rather than its size.
