Nemotron 3.5 Bolt 24GB vs Qwen3.8-27B: Both Fit in 24GB — Only One Tells You What It Cost

A neat story circulates about these two models: the N​VIDIA N​emotron 3.5 Bolt 74B-A9B 24GB fits in 24 gigabytes and Qwen3.8-27B does not. The first half is true and the second half is out of date. A​libaba published Qwen3.8-27B in August 2026 as a dense 27-billion-parameter vision-language model whose 16-bit weights run to roughly 56GB, and within days the community had produced four-bit builds of it at 16.46GB. Both models fit in 24GB today. The difference worth arguing about is that N​VIDIA designed the compression, published what it cost, and shipped it as the product — while the Q​wen side of that comparison is a build someone else made, with a bill nobody has totalled.

What each model actually is

Bolt is a mixture-of-experts hybrid: 74 billion parameters total, 9 billion active per token, Mamba-2 state-space layers interleaved with experts and attention, plus a C-RADIO v4-H vision encoder. It exists in its 24GB form because N​VIDIA ran a router-guided architecture search against a footprint budget and then quantised the experts to 2.0625 bits per element, repairing the damage with quantisation-aware distillation.

Qwen3.8-27B is dense. All 27 billion parameters are read for every token, and the model card's architecture section describes a 64-layer stack arranged as sixteen repeats of three gated linear-attention blocks followed by one gated full-attention block, with multi-token prediction trained across multiple steps and a native context of 262,144 tokens extensible to one million. It natively understands images and video, and it exposes thinking control per request through a reasoning-effort setting and a flag that preserves reasoning context across turns.

Different production philosophies, and both are shipped as open weights under permissive-enough licences — OpenMDW-1.1 for Bolt, Apache 2.0 for Qwen3.8-27B.

Two compressions, one of which is documented

This is the part of the comparison that decides things.

N​VIDIA's 24GB checkpoint is compressed by N​VIDIA, and the cost is published. The BF16 release of the same model carries a table with both columns side by side, so the price of 2-bit experts is visible line by line:

• GPQA Diamond — 86.43 at 2-bit against 88.83 at 16-bit, a loss of 2.40 points.

• Terminal-Bench 2.1 — 50.98 against 53.93, a loss of 2.95.

• BrowseComp — 66.75 against 70.00, a loss of 3.25.

• AA-LCR long-context reasoning — 64.06 against 67.38, a loss of 3.32.

• HLE without tools — 25.30 against 28.73, a loss of 3.43.

• AA-Omniscience accuracy — 16.90 against 21.80, a loss of 4.90.

And two rows that go the other way: on AA-Omniscience non-hallucination the compressed model scores 61.27 against the full-precision 56.14, and on WebArena vision the pair are level within a point. Those are vendor numbers, unreproduced, and they are still the most complete account any lab has published of what a 24GB compression did to its own model.

Now the Q​wen side. A four-bit build of Qwen3.8-27B at 16.46GB exists, was published by a third party, and comes with no accuracy measurement at all. Not a weak one — none. The same is true of the 9.83GB and 7.27GB builds sitting below it, and of the 21.98GB build above. Whatever a 27B dense vision-language model loses when its weights are squeezed to four bits, nobody has published a number for it, which means the memory arithmetic that makes the fit possible is the easy half and the capability question is entirely open.

That asymmetry is the honest version of this matchup. Bolt tells you its price. Qwen3.8-27B makes you pay one you have to estimate yourself.

The independent panel, and what it does and does not cover

Artificial Analysis measures Qwen3.8-27B at the full model level in its maximum reasoning configuration: an Intelligence Index of 34, first of 141 models in the panel; about 41 output tokens per second, which it ranks 51st of 63; and a cost of $0.50 per million input and $3.00 per million output on the provider mix it tracks, which it calls particularly expensive for an open-weights model of that size. It also notes the model is very verbose, at 200 million tokens generated during the index run.

The same panel scores Bolt at an Intelligence Index of 21, fourth of 67 models in its class, with speed and cost both unavailable because there is no public API serving it. Thirteen points is a wide gap and it is measured the same way for both, which makes it the most trustworthy single statement in this article. It also comes with a caveat that matters more than it sounds: the panel's own comparison summary calls Qwen3.8-27B notably slow and very verbose, so the model that wins on intelligence is the one you wait longer for.

On modality the two pages disagree in a way worth noting, since a spec panel is not the same as a product page. The Q​wen vendor card describes native image and video understanding; the analysis panel lists text, image and video input for Qwen3.8-27B on some views and text-and-image for Bolt. Both models see. Neither sees identically, and the difference in how third parties describe the two is a reminder to check a modality claim against the model card rather than a summary row.

The vendor table on Q​wen's side

A​libaba's model card is a long one, and its agentic rows are the reason this model is interesting next to Bolt at all. Terminal Bench 2.1 at 73.0, SWE-bench Pro at 61.7, GPQA Diamond at 89.2, HLE at 30.8, LiveCodeBench v6 at 90.3, OSWorld-Verified at 84.3 and AndroidWorld at 81.9 — all evaluated by A​libaba, in most cases inside the C​laude Code harness, at a 256,000-token context.

Those numbers are not comparable to Bolt's table. Different harnesses, different evaluation dates, different baseline selections, and in Bolt's case a baseline group chosen partly for fitting into 24GB at four-bit. The only honest statement is that Qwen3.8-27B's published agent scores are higher than Bolt's published agent scores, that both sets of numbers come from the labs that built the models, and that neither lab has run the other's benchmark suite.

One Q​wen number is worth flagging for a different reason. Its own card concedes GPQA Diamond at 89.2 while Bolt's card claims 86.43 — close enough that a harness difference could move either one past the other. That is the scale on which vendor tables settle arguments. It is not a scale on which you should settle one.

Renting one, and the latency you inherit

Bolt has no endpoint. There is no rate to quote, no provider serving it, and no way to evaluate it on your own traffic without completing N​VIDIA's deployment recipe — the pinned nightly vLLM container, the kernel package fetched at runtime, the reasoning-parser plugin, the b12x backends, the fp8 KV cache and a served concurrency of one sequence.

Qwen3.8-27B is downloadable and callable, and that combination is the reason this matchup has a practical answer at all. We serve qwen/qwen3.8-27b on our own infrastructure at $0.33 per million input tokens and $2.40 per million output. The catalogue entry lists a 262,144-token context with text, image and video input, and publishing at that rate is possible on our side because OrcaRouter passes provider list price through with no markup added — if a vendor cuts a price, the change is live here the same day rather than at the next billing cycle.

The latency, though, deserves to be quoted honestly rather than flatteringly. Over the seven days ending 16 October 2026 our own endpoint for this model reported a p50 time to first token of 10.00 seconds and a p95 of 10.00 seconds, an output speed of 515 tokens per second, and an error rate of 2.1%. The daily series explains the aggregate: 1.42 seconds on 10 October, 1.65 on the 11th, 2.83 on the 12th, then 10.00 on the 13th and the 15th, 1.91 on the 14th. Two of six days hit the ceiling of our measurement window and three did not.

That is not a nice number, and it is the number. It is also the concrete reason automatic failover exists as a routing feature rather than a slogan: a first-token wait that is fine on Monday and ten seconds on Wednesday is exactly the failure mode a fallback chain is for, and the routing DSL exists so that "try this model, fall back to that one if it degrades" is configuration rather than application code. A model you own — Bolt, or a local Q​wen build — has no such problem and no such option. It is as fast as your GPU is, every day, with no third party in the path and no second model to reach for when it stalls.

How this actually resolves

There is no winner to hand over, because the two models are not competing for the same slot.

If your constraint is a machine without a network, or work that cannot leave the device, Bolt is the answer and the published compression cost is the price of admission. Nothing about Qwen3.8-27B changes that, because you cannot download your way into a 24GB deployment that N​VIDIA has not already built and documented — you can approximate it with a four-bit build whose quality is unmeasured and whose vision projector you would have to source separately.

If your constraint is capability per dollar and the network is available, Qwen3.8-27B is ahead by thirteen points on the one index that measured both, at a published rate, under a licence your legal team has seen before. The four-bit local build is a reasonable third path, and it is the one where you should expect to do your own evaluation rather than inherit someone's table.

The part both models leave to you is the same: neither has an independent benchmark against the other, neither lab has measured the other's model, and the day-one opinions circulating about them are readings of vendor cards. Download both, run your own tasks, and keep the one that finishes. That is the only result in this comparison that will not have been published by one of the two labs that built them.

What the numbers cannot tell you is whether any of it has gone wrong in the field: the documented AI incidents and agent failures around models in this class are collected in the OrcaRouter AI Incident Archive, where each entry carries its own sources, severity, confidence grade and disputed flag.