NVIDIA Nemotron Labs Dory 1.0 4B vs Qwen 3.8: The Open-Weight 4B Against Alibaba's Three-Tier Line

Comparing N​VIDIA N​emotron Labs Dory 1.0 4B against "Qwen 3.8" is not a matchup — it is a mismatch in a specific and instructive way. Qwen 3.8 is not one model; at the open-weight end it is Qwen3.8-27B, a 27B dense multimodal model under Apache 2.0 with a 256K window, and above that it is Qwen3.8-Flash and Qwen3.8-Max, both larger, both served only over A​libaba's API. Dory is a single 4B, text-only, English-only checkpoint from N​VIDIA, published to Hugging Face in four quantisations under the OpenMDW 1.1 licence with its card stamped "GA (10/16/2026)" and no announcement behind it. So the real question is not whether a 4B can beat a 27B. It is whether an elastic-depth 4B you host yourself is a cheaper way to reach the tier of task that you would otherwise rent from the Qwen3.8 line — and the two most decision-relevant facts are that Qwen3.8-27B is open-weights, so both models run on your hardware, and that Dory's own card does not benchmark against any Qwen3.8 model at all.

What is actually being compared

Start with the Q​wen side, because "Qwen 3.8" collapses three different propositions.

Qwen3.8-27B is A​libaba's open-weight dense vision-language model, Apache 2.0, 27B parameters, 64 layers, 5,120 hidden dimension, with a native 262,144-token context that the card says is extensible toward one million. It accepts text, images and video and returns text, uses a hybrid Gated DeltaNet and Gated Attention layout across 16 repeated blocks, and ships in Transformers format with support for vLLM, SGLang and TokenSpeed. It reports 90.3 on LiveCodeBench v6, 89.2 on GPQA Diamond, 79.5 on IFBench and 73.0 on Terminal-Bench 2.1, among a long list of agentic and vision results. It has been public since mid-August 2026 and has been pulled several million times.

Qwen3.8-Flash and Qwen3.8-Max are the hosted tiers. Flash is the cost-optimised multimodal model at a million-token context; Max is A​libaba's flagship, also a million tokens, positioned against the strongest proprietary models. Neither has open weights. You reach them through an API.

N​VIDIA N​emotron Labs Dory 1.0 4B is the 4B. Dense, looped, latent-recurrent — an 8-layer input block, a 20-layer recurrent block applied a variable number of times, an 8-layer output block — with hybrid interleaved sliding-window and full attention. Text in, text out, English only, 256K context. Pre-trained on 6 trillion tokens, the first ~4.5 trillion of them natively in NVFP4, plus a 300-billion-token context extension, then SFT and multi-environment GRPO. No hosted endpoint exists for it that we can verify.

One point to get right, because it trips people up: N​VIDIA ships Dory in BF16, MXFP8, NVFP4 and Q5_K_M GGUF. The canonical repository id is the MXFP8 one, but the benchmarks in the card were measured on BF16. MXFP8 is a quantisation of the same weights, not a different model.

The spec contrast

• Parameters — Dory 1.0 4B is 4B dense, with a recurrent block applied repeatedly so effective compute per token at four loops exceeds the parameter count. Qwen3.8-27B is 27B dense with a vision tower. They are not in the same weight class and only one of them fits in a small-memory deployment.

• Modality — Qwen3.8-27B takes images and video as well as text. Dory takes text. Any workload with an image in it ends here.

• Context — both advertise 256K. Qwen3.8-27B's card says native 262,144 positions extensible to roughly a million; Dory's says up to 256K with 262,144 validated in vLLM under the shipped launch command. Qwen3.8-Flash and Max both reach one million tokens outright.

• Licence — Apache 2.0 for Qwen3.8-27B, OpenMDW 1.1 for Dory. Both commercial-friendly; Apache 2.0 is the more familiar text.

• Languages — Qwen3.8-27B is multilingual by lineage; Dory is English only with 43 programming languages in its pre-training mix.

• Compute behaviour — this is the only row where Dory has something Qwen3.8-27B structurally cannot offer. Dory's depth is a serving-time parameter: one loop for latency, two as N​VIDIA's recommended default, four when agentic accuracy matters more, all from one checkpoint, with a last-loop KV-caching mode intended to hold cache memory at roughly the one-loop size regardless of loop count. Qwen3.8-27B has fixed depth; its levers are quantisation and hardware.

• Formats — Dory ships four quantisations including a GGUF for llama.cpp. Qwen3.8-27B has a separate FP8 repo and a very large derivative ecosystem of quantisations and finetunes, which is what several million downloads buys you.

The benchmark question N​VIDIA's own table dodges

Here is the uncomfortable detail. N​VIDIA's benchmark table for Dory compares it against Qwen3.5-4B, Qwen3.5-9B, Gemma 4 26B A4B and Qwen3.6-35B-A3B. It uses the two previous Q​wen generations and never reaches for Qwen3.8. That is a defensible choice for a 4B — Qwen3.8-27B is nearly seven times its size — but it means the manufacturer has no published head-to-head against the current Q​wen generation at all, in either direction. A​libaba's Qwen3.8-27B card, for its part, benchmarks against the proprietary frontier, not against a 4B.

What Dory's card does publish, against its older comparators, is honest about where looping pays. Read across its three loop settings against Qwen3.5-4B and Qwen3.5-9B:

• SWE-bench Verified — 50.3 / 59.5 / 61.8 against 54.7 and 61.0. At four loops Dory nominally beats the 9B Q​wen on this row.

• Terminal-Bench 2.1 — 12.5 / 18.1 / 22.2 against 15 and 24. It loses to both, including the 4B.

• τ³-bench (Banking) — 11.6 / 12.89 / 14.69 against 5.03 and 7.22. Dory more than doubles the 4B Q​wen and roughly doubles the 9B.

• GPQA Diamond, no tools — 69.19 / 72.41 / 72.16 against 77.71 and 80.93. It loses to both.

• IFBench — 74.29 / 77.94 / 79 against 53.83 and 59. A large win on instruction following.

• AA-LCR — 63 / 71 / 73 against 64.7 and 68.4. Long-context retrieval improves sharply with loops and ends ahead.

• AA-Omniscience — 10.03 / 10.08 / 10.03 against 15.02 and 18.05. Stored knowledge barely moves with loop count, which is the sanity check that the rest of the table passes.

Every one of those figures is N​VIDIA's, measured on N​VIDIA's harness, with disclosed repeat counts — SciCode eight runs, AA-LCR sixteen, several benchmarks exactly once — and the card labels its Artificial-Analysis-shaped scores "internal estimates" that may deviate from official ones. None of it is reproduced.

Now put Qwen3.8-27B's card next to it. Its reported 89.2 on GPQA Diamond, 79.5 on IFBench and 73.0 on Terminal-Bench 2.1 are all above Dory's best numbers on the same benchmarks, which is exactly what you would expect from a model seven times the size with a two-generation-newer training pipeline. There is no scenario in which a 4B beats a well-executed 27B on a knowledge or reasoning benchmark, and Dory does not claim to. What Dory claims is a different thing: that at 4B it reaches roughly 9B-class agentic behaviour, and that you get to choose how much compute to spend on that scale.

Deployment: the gap is not close, and it is not about quality

Qwen3.8-27B is a normal open-weight release. Transformers format, listed support for vLLM, SGLang and TokenSpeed, an FP8 sibling, and an ecosystem that has produced well over a thousand quantisations and hundreds of finetunes since August. If your stack runs modern open-weight models, it runs this one, and the only real constraint is memory — 27B dense at BF16 is not a small-machine model, but the quantised variants exist precisely for that.

Dory is the opposite. vLLM support is under review in an open pull request; until it merges, N​VIDIA tells you to install vLLM from that branch with the precompiled-kernel flag and then launch with overrides configuring the recurrent loop count and the KV-cache mode. llama.cpp support is in a second open pull request with a pinned commit and ARM64 macOS build instructions. There is no hosted endpoint for Dory anywhere we can verify, and the download counters on all four repositories were at zero when this was written. Published speed figures are limited to quantisation speedups on a DGX Spark — 1.69× and 1.84× for MXFP8 over BF16, 2.17× and 2.44× for NVFP4, depending on the input/output length mix — not throughput.

Set against that, the memory claim in Dory's card is the one piece of engineering that could justify the inconvenience. Standard recurrent looping stores a separate KV cache per loop, so memory grows with every pass. Dory's last-loop KV-caching mode keeps one cache per recurrent attention layer and overwrites it each pass, holding the footprint at roughly the one-loop size for any loop count, with accuracy that N​VIDIA says stays close to the per-loop mode. If that holds under someone else's measurement, a 4B with a 256K window and four loops becomes practical on hardware that would not otherwise afford it. Nobody outside N​VIDIA has checked.

Where OrcaRouter actually sits in this comparison

This is the rare comparison where we can be specific, because the two sides are on different footing here. The Q​wen side is on our catalogue: qwen/qwen3.8-27b is self-hosted on OrcaRouter's own infrastructure, and qwen/qwen3.8-flash and qwen/qwen3.8-max are reachable as hosted endpoints, all through one API key at provider list price with 0% markup. The N​VIDIA side is not — N​VIDIA N​emotron Labs Dory 1.0 4B is not on the catalogue and has no hosted endpoint anywhere we can confirm.

That asymmetry is genuinely useful rather than awkward. The argument for a small self-hosted model is a cost argument, and a cost argument needs the rental price it is arguing against. The three Qwen3.8 tiers price very differently — 27B at $0.33 per million input tokens and $2.40 per million output, Flash at $0.15 and $0.47, Max at $2.00 and $6.00 — so the honest version of "can a 4B replace this" has three different answers depending on which tier you are actually using today. A single OrcaRouter key reaches all three, so pricing that question does not require three accounts. Because we pass provider list price through with no markup, a Q​wen price change lands here the same day. And when you do pick a tier, automatic failover across providers and the routing DSL — which lets you compose models behind one call — mean the switch from a hosted Q​wen to a self-hosted Dory, or back, is a config change at one call site rather than a refactor. We do not serve Dory; we are the cheap way to measure what it would be replacing.

The verdict

If your workload is multimodal, multilingual, or knowledge-heavy — anything where the model needs to know things rather than do things — Qwen3.8-27B is the open-weight answer and the comparison is not competitive. It is seven times the size, two generations newer, backed by a large derivative ecosystem, runs on engines that already support it, and its own card claims numbers Dory does not approach on GPQA or instruction benchmarks. Rent Flash or Max instead if you would rather not host a 27B at all.

Dory 1.0 4B is a bet on a different axis. It is worth the build-from-a-branch friction only if three things are true for you: the task is text-only and English-only, it is agentic rather than knowledge-driven, and you would genuinely use a compute dial — running one loop under load and four when accuracy matters, from one checkpoint, without a second download. Its card claims that bet pays on τ³-Bench and instruction following at 4B, and it discloses enough about how those numbers were produced to let you discount them appropriately.

What would settle it is the test nobody has run: Dory 1.0 4B against Qwen3.8-27B on a shared harness, at the same quantisation, on an agentic task set. N​VIDIA's table skips the current Q​wen generation entirely and A​libaba's table does not look down a weight class. Until someone publishes that comparison, the accurate position is that one of these models is a proven, supported, widely-deployed 27B and the other is a 4B repository that appeared overnight with an interesting idea in it — and the only person who can say whether the idea is worth the build is you, on your task set.

None of that is hypothetical: the documented AI failures and agent incidents behind releases like this one are collected in the OrcaRouter AI Incident Archive, where each entry carries its own sources, severity, confidence grade and disputed flag.