Nemotron 3.5 Bolt 74B vs DeepSeek V4 Pro: Both Open, One Fits in 24GB

NVIDIA's Nemotron 3.5 Bolt 74B-A9B 24GB and DeepSeek V4 Pro are on the same side of the open-versus-closed line and nowhere near each other on the hardware line. Both publish weights you can download, both permit commercial use, and both can be served without asking anyone's permission. One of them is 74 billion parameters with 9 billion active, quantised to fit in 24 gigabytes so it runs on a laptop-class machine next to your other applications. The other is 1.6 trillion parameters with 49 billion active, and the phrase "deployment footprint" does not mean what it means for the first one.

That difference is more consequential than the index scores, and the index scores are already 15 points apart. Below is the comparison a team actually has to run: not which model is better in the abstract, but which of these two open-weight products you can physically operate, at what quality, for what money.

Two very different shapes of "open"

Worth stating plainly before any of it, though, because the two open-weight halves did not arrive the same way. DeepSeek published V4 Pro as a product, with a card and a rate card. NVIDIA has not announced Nemotron 3.5 Bolt 74B-A9B 24GB at all. The files simply appeared: the 24GB quantised repository on 7 October 2026, its GGUF conversion hours later, a BF16 checkpoint on 9 October, the LiLiCorr speculative drafter on 12 October. Sixteen days of silence on NVIDIA's own channels is the state of things, so the verifiable half of this comparison is what a repository can prove — licence, parameter counts, file formats, the serving recipe on the card — and the unconfirmed half is everything a launch post would have carried. The throughput and agentic figures below are NVIDIA's own and are marked as such, because there is no third party except one index to check them against.

Start with the licence and the checkpoint, because that is where an open-weight decision is really made.

• Licence — Nemotron 3.5 Bolt 74B-A9B 24GB: OpenMDW-1.1, a permissive model-weights licence with an explicit commercial-use carve-in; NVIDIA's card states the model is ready for commercial or non-commercial use. DeepSeek V4 Pro: MIT, which is about as unrestricted as a weights licence gets.

• What is actually publishable — Nemotron 3.5 Bolt 74B-A9B 24GB: the 24GB IQ2 build, a BF16 checkpoint before quantisation, GGUF builds for llama.cpp, and a separate LiLiCorr speculative drafter. DeepSeek V4 Pro: the full-weight checkpoint and its derivatives.

• Where the checkpoint came from — Nemotron 3.5 Bolt 74B-A9B 24GB: compressed from NVIDIA's Nemotron 3.5 Super VL 120B-A12B parent using Star Elastic, a router-guided architecture search, then quantised with quantisation-aware distillation. It is a distilled child, not a from-scratch model, and NVIDIA's own research page argues explicitly that compressing a large model into a fixed budget beats training a small one from scratch. DeepSeek V4 Pro: a natively trained frontier-scale mixture-of-experts.

That third bullet is the one that decides most evaluations you will read. Any benchmark comparison between these two is a comparison between a design philosophy of aggressive compression and a design philosophy of scale, and the two do not converge on the same workloads.

What each one measures at

Artificial Analysis has run both, on the same current revision of its index, which is the only reason a sentence like this can be written at all.

• Intelligence Index — Nemotron 3.5 Bolt 74B-A9B 24GB 21, fourth of 67 models in its open-weight size class. DeepSeek V4 Pro 36, ninth of 116, measured on the 0813 build in the max configuration.

• Cost per Intelligence Index task — Nemotron 3.5 Bolt 74B-A9B 24GB: not applicable, because the panel records no public API for it and there is no metered price to divide by. DeepSeek V4 Pro: 0.67 dollars measured on the panel's own weighted basis.

• Output speed — Nemotron 3.5 Bolt 74B-A9B 24GB: NVIDIA reports 83.99 tokens per second on AIPerf SPEED-Bench-Coding with its drafter, on a Surface Laptop Ultra. DeepSeek V4 Pro: 96.3 tokens per second measured by Artificial Analysis against the hosted API. The two are not the same measurement and should not be subtracted from each other; what they establish is that both are fast enough that speed is not your constraint.

• Verbosity — Nemotron 3.5 Bolt 74B-A9B 24GB generated 230 million tokens during the panel's index run, against a median of 100 million for its class — the panel's own note calls it very verbose. DeepSeek V4 Pro generated 160 million against a median of 130 million. On a metered API that difference is a line item; on a local box it is electricity and wall-clock.

The 15-point index gap is real and is not closed by the fact that the smaller model is cheaper to run. What closes it, in the workloads where it closes at all, is the vendor's own benchmark set — and that set is worth reading with the right suspicion. NVIDIA's card puts the Bolt model at Terminal-Bench 2.1 of 50.98 and BrowseComp of 66.75, ahead of every comparator column printed beside it. Those are NVIDIA's numbers on NVIDIA's card, and DeepSeek is not one of the columns.

The deployment diagram is the actual decision

This is where DeepSeek V4 Pro and Nemotron 3.5 Bolt 74B-A9B 24GB stop being substitutes for one another.

• Room to run. A 1.6-trillion-parameter model at 49 billion active needs a serving cluster. Even heavily quantised, the checkpoint alone is an order of magnitude past what a single workstation can hold, and the active-parameter count means every token touches a substantial slice of memory bandwidth regardless of the quantisation scheme. The Bolt model was designed against a stated ceiling — NVIDIA's architecture search targets a footprint below 24GB at batch size one and 128K context — and the search is run with the downstream quantisation already accounted for.

• Minimum viable serving stack. Nemotron 3.5 Bolt 74B-A9B 24GB wants a specific vLLM nightly with the v2 model runner enabled, a downloaded reasoning-parser plugin, b12x kernels, and explicit MoE, Mamba and attention backend flags; llama.cpp covers the GGUF path. DeepSeek V4 Pro is served by the usual inference stacks at distributed scale, which is a different kind of project with a different kind of on-call rotation.

• Inputs. The Bolt model takes text, images and video, with the reference recipe capping video frames and applying a pruning rate. DeepSeek V4 Pro's published modality list is text in, text out — its mixture-of-experts scale is spent on reasoning, not perception.

• Context. 262,144 tokens for the Bolt model against roughly a million for DeepSeek V4 Pro. If your workload is a repository, a contract set or a long agent transcript, that is a real four-times difference and it is in DeepSeek's favour.

Put plainly: if you have a datacenter and a workload that needs frontier reasoning, DeepSeek V4 Pro is the better model and the index says so. If you have a laptop, a Spark box or a workstation, and the requirement is that the model runs locally alongside everything else, Nemotron 3.5 Bolt 74B-A9B 24GB is not a compromise on that requirement — it is the only one of the two that meets it.

What the smaller model buys you beyond the disk

Three things that do not show up in an index score and do show up in a deployment review.

• It runs with the network cable out. A 24GB checkpoint on local storage has no availability dependency, no rate limit and no provider incident. For an agent that runs on a machine you own, that is the whole design goal and the reason NVIDIA compressed a 120B parent instead of shipping something at full precision.

• Speculation is cheap and safe here. The LiLiCorr drafter adds under a gigabyte on disk and NVIDIA reports mean acceptance lengths of 4.68 greedy and 4.25 probabilistic at seven drafted tokens and concurrency 64. Because the target verifies every draft, a bad guess costs speed rather than accuracy — the output distribution is preserved. That property is what makes shipping a speculative path reasonable on a local box, and it is documented on the model card.

• The whole stack is inspectable. Weights, quantisation config, drafter, serving recipe and the technical report are all public. When the model misbehaves you can read the config rather than filing a ticket.

Costing the two options honestly

The price comparison is not symmetric and pretending it is would be the easiest mistake on this page.

• Nemotron 3.5 Bolt 74B-A9B 24GB has no per-token price because it has no metered endpoint. Its cost is hardware you buy or already own, plus power, plus the engineering hours to stand up the serving recipe. Once that is paid, marginal cost per token is close to zero and stays there.

• DeepSeek V4 Pro is metered. On OrcaRouter it is routed at DeepSeek's published list, 0.66 dollars per million input tokens and 1.98 dollars per million output tokens, with no markup added on top — so a vendor price move is live on our side the same day. For a workload in the millions of tokens a day, that is a budget line, not a rounding error.

• The break-even is not a token count, it is a utilisation profile. A local model that sits idle is pure capex. A hosted model that runs one request a week is almost free. The crossover happens when the tokens you would meter exceed the amortised cost of the box, and the only way to know where that is for your traffic is to measure a week of it.

There is also a third option that costs almost nothing to try, which is that OrcaRouter routes DeepSeek V4 Pro on one key alongside 200-plus other models, with automatic failover across providers. Before committing a workstation to a 24GB local build, comparing that build's output against a metered frontier model on your own prompt set costs a few dollars and settles the quality question empirically instead of on the strength of an index table.

The version of this decision nobody regrets

Run both, in the pattern each is good at. DeepSeek V4 Pro for the requests where reasoning depth or a million-token context is the point, and where the metered cost is visible and acceptable. Nemotron 3.5 Bolt 74B-A9B 24GB for the high-volume, private, multimodal or offline traffic, where the model being local is the requirement rather than a cost optimisation.

The comparison people keep trying to run — one index number against another — is the least useful frame here, because the two products were not built to compete. One was built to be as good as possible without a footprint constraint. The other was built to be as good as possible inside 24 gigabytes. Both are open, both are legitimate, and that is exactly why the choice is yours to make rather than a vendor's to make for you.

Comparisons like this one are also where the failure modes show up: the documented AI incidents and agent failures behind 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.