Nemotron 3.5 Bolt 74B vs Qwen3.8-Max: The Widest Gap on the Board

Every matchup in the N​emotron 3.5 Bolt family has a gap in it. This one has the widest. N​VIDIA N​emotron 3.5 Bolt 74B-A9B 24GB scores 21.4 on the Artificial Analysis Intelligence Index; Qwen3.8-Max, A​libaba's flagship and the highest-scoring model in this series of comparisons, scores 45.4 on the same independently run board. That is 24 points, more than the entire index score of the smaller model, and it is the honest starting point for anyone asking whether a 24GB local agent can stand in for the frontier.

It cannot. What makes the pair worth writing about is that the two models win on entirely different rows, and that N​VIDIA's own benchmark table compares Bolt to a Q​wen model that is not this one.

Two models with nothing in common except multimodality

N​VIDIA N​emotron 3.5 Bolt 74B-A9B 24GB is a 74-billion-parameter vision-language model compressed to a 2-bit checkpoint that fits in 24 gigabytes of memory. Its weights went up on Hugging Face on 7 October 2026; N​VIDIA named it in a launch post on 15 October. It is built for local agentic work — tool calling, web browsing, terminal use, reading a screen — and it is sold as a download rather than a service, with no published metered price anywhere.

Qwen3.8-Max is the opposite construction. A​libaba's top-tier model is a sparse mixture-of-experts that the company positions directly against the strongest closed models of its release window, served through three wire formats — an O​penAI-compatible surface, an A​nthropic-compatible surface and A​libaba's own DashScope — across regions in Beijing, Singapore, Tokyo, Frankfurt and Virginia. The hosted model launched on 3 August 2026 with image and video input, a million-token context and a thinking toggle; the September 2 snapshot is the version currently scored on the leaderboards. The open-weight sibling in the same family, Qwen3.8-2.4T-A95B, is a separate release and not the model measured here.

One is a file. One is a service. The comparison still has a use, because the service publishes a price and the file does not, and the difference between them is what a local deployment is actually buying.

• Intelligence Index — N​VIDIA N​emotron 3.5 Bolt 74B-A9B 24GB: 21.4. Qwen3.8-Max: 45.4. Both independently measured by Artificial Analysis.

• Context — 262,144 tokens for the N​VIDIA checkpoint against roughly 983,000 for Qwen3.8-Max, with a 131,072-token maximum output on the A​libaba side.

• Parameters — 74B total and 9B active for Bolt; A​libaba's flagship tier is a sparse mixture-of-experts of a different order entirely, activating roughly 95B parameters per token in its open-weight form.

• Inputs — text, images and video in, text out for both. This is the one dimension on which the two are genuinely alike.

• Price — Qwen3.8-Max: $2.00 per million input tokens, $6.00 per million output, $0.25 for cached input. Bolt: nothing, because there is no endpoint to buy.

• Cost per Intelligence Index task — $5.41 for Qwen3.8-Max, measured. No figure for Bolt, for the same reason.

• Speed — Qwen3.8-Max: 39.8 output tokens per second with 3.06 seconds to first token and a median 50.3 seconds of reasoning time per response, all independently measured. Bolt: no independent throughput measurement exists.

The rows, where the 24 points come from

Artificial Analysis ran both. Reading the row-by-row detail is more useful than the composite, because the composite hides the fact that neither model is uniformly ahead.

• Terminal-Bench 2.1 — Qwen3.8-Max 88.8 percent, independently measured. Bolt has no independent entry on this benchmark; N​VIDIA reports 50.98 on its model card.

• GPQA Diamond — Qwen3.8-Max 92.8 percent independently. Bolt's 86.43 is N​VIDIA's own figure, and the panel has no independent number for it.

• Humanity's Last Exam, no tools — Qwen3.8-Max 43.1 percent against Bolt's 24.7 percent, both independently measured.

• Long-context retrieval (AA-LCR) — 80.3 percent against 69.7 percent, both measured.

• Scientific coding (SciCode) — 52.1 percent against 39.7 percent, both measured.

• Multimodal reasoning (MMMU-Pro) — Qwen3.8-Max 82.8 percent, independently measured. Bolt has no independent entry, which is a notable omission for a model whose main selling point is that it can see.

• Occupational deliverables (GDPval-AA) — Qwen3.8-Max 0.583 normalised against Bolt's 0.240, both measured.

• Agentic and IT operations — Qwen3.8-Max posts 42.4 percent on APEX-Agents, 40.3 percent on ITBench-SRE and 45.0 percent on an analyst-agent evaluation. Bolt has no independent scores in any of those three columns.

• Non-hallucination (AA-Omniscience) — Qwen3.8-Max 71.2 percent against Bolt's 56.3 percent, with accuracy at 31.7 percent against 17.1 percent. Unlike the Bolt-versus-Ultra comparison, where the compressed model wins the calibration row, here the flagship takes both.

• Cost per index task — $5.41 for Qwen3.8-Max. This is the highest cost-per-task figure in the family of comparisons around this model, and it is the number that makes the local option interesting even at a 24-point deficit.

What $5.41 a task buys, and what 24GB buys instead

Qwen3.8-Max's economics are unusual for a frontier model: the per-token price of $2.00 in and $6.00 out is moderate, but the model spends a long time thinking. Median reasoning time on the panel's runs is 50.3 seconds, and end-to-end response time averages 65.9 seconds. A task that takes the model a minute of internal deliberation generates a lot of output tokens, and cost per index task is the metric that exposes it. That is not a criticism of the model — deep reasoning is the product — but it is the reason a high-volume agent loop gets expensive quickly on this tier.

That is precisely the workload Bolt is built for. A local checkpoint has no per-token meter at all: the marginal cost of a tool call is electricity, and the fixed cost is a GPU. For an always-on agent doing thousands of small procedural steps — reading a screen, calling an endpoint, editing a file — the comparison is not 21.4 against 45.4, it is a constant against a growing invoice. N​VIDIA's own framing is that Bolt exists so that everyday agentic tasks can run locally without sending the workload to the cloud, and it reports 50.86 on WebArena vision and 66.75 on BrowseComp for exactly those tasks.

The catch is that the fixed cost is real and the latency is unknown. N​VIDIA's reference serving recipe is multi-step: a pinned vLLM nightly build, the b12x kernel package, a separately downloaded reasoning-parser plugin, explicit Mamba and attention backends, and a speculative-decoding configuration using the LiLiCorr drafter with seven draft tokens. Throughput is published by N​VIDIA at 83.99 tokens per second on AIPerf SPEED-Bench-Coding on a laptop-class machine — a vendor number on a vendor benchmark, with no independent run to check it against.

The baseline N​VIDIA chose

There is one more thing worth noticing, and it is a reading lesson rather than a technical one. The benchmark table on Bolt's model card includes a Q​wen model as a comparison point — and it is Qwen3.6-35B-A3B, a much smaller model than the flagship measured on this page. N​VIDIA's table shows Bolt winning on Terminal-Bench 2.1, GPQA Diamond, BrowseComp, GDPVal, Tau3 Banking, SciCode, AA-LCR and MCP-Atlas against that model, while losing on AA-Omniscience accuracy, AA-Omniscience non-hallucination and PinchBench.

Those comparisons may well be correct. But a reader who takes "Bolt beats Q​wen" from the table is comparing a 74B edge model to a 35B dense-class MoE, not to A​libaba's frontier tier. Against Qwen3.8-Max the independent record runs the other way on nearly every shared row, and the gap widens rather than closes on the hardest benchmarks. When a vendor picks its own baselines, the honest move is to check what the baseline is.

How to decide

Choose Qwen3.8-Max when correctness is the constraint. If the task is multi-step analysis, hard multimodal reasoning, agentic work with real consequences, or anything measured by the benchmarks on which it scores 24 points ahead, the local option is not a candidate — and at $2.00 and $6.00 per million tokens, it is priced as infrastructure rather than as a luxury. It is also the model to pick when the workflow needs structured outputs, native tool calling and a million-token context in one call, all of which A​libaba documents as supported.

Choose Bolt when the constraint is where the work happens. Privacy, offline operation, latency that does not depend on a network, and a per-call cost that does not exist are four separate arguments, and each one is sufficient on its own for some teams. The model is competent at procedures and weak at recall; point it at the former.

For the hosted half, Qwen3.8-Max is reachable through OrcaRouter as qwen/qwen3.8-max/ on the same key as 200-plus other models, at Alibaba's provider list price with no markup added — which means a vendor price change lands on our side the same day rather than at the end of a billing cycle. That matters more on this tier than on a cheap one, because $5.41 per index task is a number a finance team will want to re-check monthly. Where routing changes the shape of the decision is the volume problem: a workload that is 90 percent small procedural steps and 10 percent hard reasoning can send the 10 percent here and keep the rest local, with automatic failover across providers so the hosted tier degrades into a retry rather than an outage.

We host no N​VIDIA model and have no Bolt endpoint — the N​emotron family is not in our catalogue, and the local half of this comparison is a hardware question that no routing layer touches.

The unresolved number

What nobody has published is the measurement that would make this comparison complete: an independent latency and throughput run on the 24GB checkpoint at batch size one, on a machine a person can buy. If a 2-bit 74B checkpoint turns out to run interactive agent steps at a usable speed, the 24-point gap becomes a gap you can route around — local for procedure, hosted for reasoning. If it does not, the local option is a batch tool and the comparison collapses to price. Either way, the answer is one publication away, and it is the thing to look for before planning a deployment around this model.

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.