Nemotron 3.5 Bolt vs Qwen 3.8 Max: Both of These Models Watch Video

There are not many model pairings where the surprising part is a capability they share. Both NVIDIA Nemotron 3.5 Bolt 74B-A9B 24GB and Qwen 3.8 Max accept video as input, and both return text. Qwen 3.8 Max, published by Alibaba on 3 August 2026, is a hosted multimodal model with a 1,000,000-token context, served through OpenAI-compatible, Anthropic-compatible and native DashScope interfaces from five regions, priced at $2.00 per million input tokens and $6.00 per million output tokens with cached reads at $0.25. Nemotron 3.5 Bolt, released by NVIDIA on 15 October 2026, is a 24GB open-weight file with a C-RADIO v4-H vision encoder, a 262,144-token window, and no price at all.

Two models that can both be handed a screen recording, separated by nothing except where they run and what that costs. That is a cleaner comparison than most head-to-heads get, and it resolves quickly once you stop asking which is smarter.

Two multimodal stacks, one memory envelope

Bolt's multimodality is inherited rather than built for it. NVIDIA compressed its 120B-A12B Nemotron 3.5 Super VL model down to 74B-A9B with a router-guided search it calls Star Elastic, then took the result to 2.0625 bits per expert weight in an IQ2_XXS format with quantization-aware distillation. The vision tower — a C-RADIO v4-H encoder — came along with the compression, and the reasoning-mode switch is the same one the Super VL client exposes: an enable_thinking flag in the chat template. Total parameters 74B, active per token 9B, backbone a Mamba-2 and Transformer hybrid the vendor calls LatentMoE, licence OpenMDW-1.1 for commercial and non-commercial use.

Qwen 3.8 Max is a different kind of artifact. It is Alibaba's highest-capability tier, natively multimodal, and it exists only as an endpoint. There is no checkpoint to download, no quantised build to run on a workstation, and no version of it whose marginal cost is the power bill. What it has instead is reach: three wire formats, so an existing Anthropic SDK or a DashScope integration works without a rewrite, and five served regions across Asia, Europe and North America that make the latency question a geography question as much as an engineering one.

The practical consequence is that Bolt's multimodal behaviour is bounded by the same 24GB that holds the KV cache, the conversation history and the tool results. Qwen 3.8 Max's is bounded by whatever the endpoint provisions, which is not your problem.

What you pay, and to whom

Qwen 3.8 Max is metered, and the rates are low for what it is. A 20,000-token prompt with a 3,000-token answer costs $0.04 in and $0.18 out, about $0.22 a call — and cached reads at $0.25 per million mean a stable system prompt or a repeated document is close to free on the second pass. Its p50 time to first token on our catalogue is 3.22 seconds with a p95 of 10.00 seconds, and it carried 41.3 million tokens of routed traffic across the last seven days.

Nemotron 3.5 Bolt has no rate card because there is nothing to meter. No hosted endpoint, no vendor API, no third-party platform serving it that we can point you at. Artificial Analysis lists both its speed and its cost per Intelligence Index task as N/A for exactly that reason. If you run it, the cost is a GPU, a power draw and an amortisation period, and the marginal price of the next million tokens is zero. For a nightly job that reads the same corpus every day, that is not a discount — it is a different accounting model, and the break-even point depends entirely on how often you run it.

The window, and the ability to use it

Context is the axis where these two look closest on paper and separate fastest in practice.

• Declared window — Qwen 3.8 Max 1,000,000 tokens against Nemotron 3.5 Bolt 262,144. Roughly four to one, and the ratio is what decides whether a whole-codebase or whole-archive task is even possible in one request.

• Recall inside the window — Artificial Analysis records a long-context recall figure of 80.3 for Qwen 3.8 Max. There is no equivalent independent number for Bolt, and its window is smaller to begin with, which makes this a comparison between a measured capability and an unmeasured one.

• What else shares the space — Bolt's 262,144 tokens live inside the same 24GB as the weights, the video frames and the KV cache. A long video prompt and a long conversation compete for the same memory. Qwen 3.8 Max's window is provisioned by the operator, so the constraint is the price rather than the hardware.

• Reasoning control — Qwen 3.8 Max exposes an enable_thinking toggle, or reasoning.effort on the Responses API. Bolt's equivalent is a chat-template switch in the same family of flags, inherited from the Super VL client. Neither is unusual; both let you stop paying for deliberation on routine calls.

Measured against unmeasured

Artificial Analysis scores Qwen 3.8 Max at 45.4 on its Intelligence Index, fifteenth of 145 models measured — inside the top ten percent without being at the top. Its supporting rows are strong where an agentic workload needs them: an AA Coding Index of 76.2 at ninth of 138, 88.8 on Terminal-Bench 2.1, and 92.8 on GPQA Diamond. Its comparative weakness is that 80.3 long-context recall, which is respectable rather than exceptional for a million-token model.

Bolt's numbers are narrower in scope. It scores 21 on the same index, fourth of 67 in its footprint-filtered class against a class median of 8, and Artificial Analysis describes it as among the leading open-weight models of similar size. Its measured speed and cost are blank. Everything else — the Terminal-Bench 2.1 score of 50.98, GPQA Diamond 86.43, BrowseComp 66.75, WebArena vision 50.86, SciCode 41.02 — comes from NVIDIA's own model card, where every baseline is another model a third the size. Those rows are not dishonest; they are answering a question about efficiency, not about frontier capability, and reading them as a contest against a hosted flagship misreads the document.

The speed gap deserves the same care. Artificial Analysis cannot measure Bolt's throughput because there is nothing to call. NVIDIA reports a design target of at least 80 tokens per second of decoding, 84 on the AIPerf SPEED-Bench on RTX Spark, and 83.99 in a llama.cpp run on a Surface Laptop Ultra with 128GB, all of it unreproduced. Qwen 3.8 Max's 3.22 seconds to first token is measured and public. Plan against the measured number and treat the other as an aspiration.

The routing question, which is not symmetrical here

Qwen 3.8 Max is on our catalogue as qwen/qwen3.8-max, routed at Alibaba's provider list price with 0% markup added — $2.00 and $6.00 per million tokens, the same numbers on our page as on the vendor's, which means a vendor-side rate change reaches our users the same day rather than at a contract renewal. It sits behind the same key as 200-plus other models, with automatic failover across upstream providers when one degrades, and a routing DSL for composing several models into a single request.

Nemotron 3.5 Bolt is not on our catalogue, and we do not serve it. There is no Bolt endpoint, no identifier to point a request at, and nothing in this piece is an offer to host NVIDIA's checkpoint. The honest shape of the routing advice here is a split rather than a choice: the hosted multimodal model handles the long-window requests, the video corpora too large for 24GB of shared memory, and anything that needs to be callable from anywhere; the local file handles the high-frequency passes where the data should not leave the building and volume would make per-token billing hurt. Model fusion is the variant worth knowing if the expensive calls are the consequential ones — a panel of models answers the same prompt together and you take the agreement, which is a cheap way to buy confidence on a decision you cannot easily reverse.

How this one resolves

The overlap is video, and it is the reason these two are worth comparing at all — most local models at this footprint cannot see at all, and most hosted models cannot be handed a screen recording without a frame-extraction step that loses the thing you were trying to show it.

Past that overlap, they diverge on every axis. Qwen 3.8 Max wins the wide window, the measured latency, the richer input set and, by a wide margin, the independently scored intelligence. Nemotron 3.5 Bolt wins the only axis that no hosted model can contest: it runs where your data already is, whether or not there is a network, at a marginal cost of zero.

The decision usually is not a preference. If the video never leaves the building, the hosted option was never available and the comparison collapses to a single candidate. If the corpus is a million tokens of text, the local option was never available for the same reason in reverse. Where it is genuinely close — modest volumes, data that could legally travel, work that is mostly short tool calls with occasional long reads — the tiebreaker is the one number missing from this comparison: how fast the 24GB checkpoint actually runs on your hardware, which nobody outside NVIDIA has yet measured.