
Qwen4-Max vs Nemotron 3.5 Bolt 74B: One Number Is Missing From This Comparison
- OrcaNEWOrcaCyber: Hy4 Flash Abliterated / Uncensored2026-11-18$0.07 / $0.28 per 1M tokens · 184 tok/s
- tencentNEWTencent: Hy4 Flash2026-11-16$0.02 / $0.07 per 1M tokens · 215 tok/s
- GoogleNEWGoogle: Gemini 4 Carbon2026-11-10$2.00 / $10.00 per 1M tokens · 87 tok/s
- AnthropicNEWAnthropic: Claude Fable 5.22026-11-09$8.00 / $40.00 per 1M tokens · 97 tok/s
- orcaNEWOrca: OrcaC2 1.02026-11-079 tok/s
- OrcaOrcaCyber: GLM 5.3 Abliterated / Uncensored2026-11-04$3.00 / $5.00 per 1M tokens · 90 tok/s
- OrcaOrcaCyber: Hy4 Abliterated / Uncensored2026-11-04$0.80 / $1.60 per 1M tokens · 131 tok/s
- OrcaOrcaCyber: Kimi K3.1 Pro Abliterated / Uncensored2026-11-04$2.00 / $5.00 per 1M tokens · 76 tok/s
- tencentTencent: Hy42026-11-04$0.83 / $2.50 per 1M tokens · 109 tok/s
- OrcaOrcaCyber: MiniMax M3.1 Flash Abliterated / Uncensored2026-10-29$0.80 / $1.60 per 1M tokens · 114 tok/s
- OrcaOrcaCyber: GLM 5.3 Flash Abliterated / Uncensored2026-10-27$0.80 / $1.60 per 1M tokens · 68 tok/s
- OrcaOrcaCyber: Qwen3.8 Flash Next Abliterated / Uncensored2026-10-26$0.80 / $1.60 per 1M tokens · 94 tok/s
- OrcaOrcaCyber: MiMo V2.6 Flash Abliterated / Uncensored2026-10-26$0.80 / $1.60 per 1M tokens · 41 tok/s
- MoonshotMoonshotAI: Kimi K3.1 Pro2026-10-2447Intelligence
- OrcaOrcaCyber: DeepSeek V4.1 Flash Abliterated / Uncensored2026-10-23$0.80 / $1.60 per 1M tokens · 209 tok/s
- minimaxMiniMax: M3.1 Flash2026-10-1929Intelligence
- OrcaOrca: OrcaCyber Ultra 1.02026-10-17$5.00 / $10.00 per 1M tokens · 111 tok/s
- deepseekDeepSeek: DeepSeek V4.1 Flash 10172026-10-1733Intelligence
- DeepSeekDeepSeek: DeepSeek V4.1 Flash 09102026-10-17$0.15 / $0.60 per 1M tokens · 311 tok/s
- DeepSeekDeepSeek: DeepSeek V4.1 Flash2026-10-1740Intelligence
Put Qwen4-Max and NVIDIA Nemotron 3.5 Bolt 74B-A9B 24GB side by side and the first thing you notice is that the comparison sites have nothing to compare. BenchLM says it plainly — no shared benchmark results yet — and then invites you to look at price instead. The reason is structural rather than temporary: Qwen4-Max scores 43 on the Artificial Analysis Intelligence Index v5.0.1, and Nemotron 3.5 Bolt 74B-A9B 24GB has no Intelligence Index score at all. Not a low one. N/A, on a board that has been running models since before either shipped, because the model has no publicly available API and the independent harness has nothing to call.
That single absence reorganises everything else about the matchup. You cannot rank a 74-billion-parameter checkpoint that fits in 24 gigabytes against a 2.4-trillion-parameter hosted model by putting their numbers in adjacent cells, because only one of them has numbers. What you can do — and what the aggregator pages do not do — is separate the claims each vendor makes, note who measured them, and work out which column the decision actually turns on. That turns out to be less about capability than about whether you are buying a file or renting a service.
One is a file, the other is a meter
NVIDIA Nemotron 3.5 Bolt 74B-A9B 24GB shipped on 15 October 2026 and lives on Hugging Face as downloadable weights under OpenMDW-1.1, a licence that permits commercial use. It is a hybrid Mamba-Transformer mixture-of-experts: 74 billion parameters total, 9 billion active per token. The "24GB" in the name is the point of the whole exercise. NVIDIA built it by compressing its 120B-A12B parent with a router-guided search framework it calls Star Elastic, cutting expert top-K from 22 to 9, then applying quantization-aware distillation so the expert layers sit at 2.0625 bits. A 1.13-billion-parameter speculative drafter rides along, with mean acceptance lengths of 4.68 under greedy decoding and 4.25 under sampling. The recommended runtime is vLLM on Ampere, Hopper or Blackwell, at a tensor-parallel size of one.
Qwen4-Max is the opposite construction, and not only because of the size. Alibaba's fourth-generation flagship went generally available on 16 November 2026 at $1.50 per million input tokens and $4.50 per million output in Singapore, with a 1,000,000-token context window, native image and video input, and three configurable reasoning levels. No checkpoint is published. Fine-tuning, model deployment and batch inference are all listed as unsupported in the Model Studio documentation. Every endpoint serving it is serving Alibaba's deployment.
So the two models are not competing for the same slot in your architecture. One is something you run; one is something you call. Any comparison that does not start there is measuring the wrong thing.
The vendor table compares Bolt to a Qwen that is not this one
NVIDIA's own benchmark table is the best-documented set of numbers on either side, and it is worth reading carefully because of what it does not contain. Its comparison columns are Qwen3.6-35B-A3B in BF16, Gemma 4 26B-A4B-IT in BF16, and Nemotron 3.5 Lightning 30B-A3B. Qwen4-Max is not in the table. The Qwen the table measures is a 35-billion-parameter open-weight model two generations back, and the reason is obvious once stated: the comparison set is defined by memory footprint, and a hosted 2.4-trillion-parameter API has no footprint to compare.
What the table does establish, all of it vendor-reported and unreproduced, is that a 24GB file is competitive with much larger local options on agentic work. On the rows NVIDIA chose:
• BrowseComp — Nemotron 3.5 Bolt 74B 66.75, against 48.74 for Qwen3.6-35B-A3B and 26.30 for Gemma 4 26B-A4B
• Terminal-Bench 2.1 — 50.98, against 44.38 and 37.22
• WebArena in vision mode — 50.86, against 36.60 and 38.20
• GPQA Diamond — 86.43, against 83.40 and 79.61
• HLE without tools — 25.30, against 19.56 and 17.42
• GDPVal-AA-V2 — 28.47, against 25.75 and 15.33
• MCP-Atlas — 52.20, against 51.20 and 48.4
• Tau3 Banking — 17.01, against 10.52 and 14.02

The two rows where it loses are worth naming, because vendor tables rarely include them. On AA-Omniscience accuracy, Gemma 4 26B-A4B scores 22.17 against Bolt's 16.90; on AA-Omniscience non-hallucination, Nemotron 3.5 Lightning leads at 68.81 against Bolt's 61.27. Bolt's weakest disclosed property is knowing when it does not know.
The footnote matters as much as the rows. NVIDIA records MCP-Atlas at 51.20 for Qwen3.6-35B-A3B while noting that Qwen itself reports 62.8 on the same benchmark, measured on the 500 public tasks. Two labs, one benchmark, a gap of eleven and a half points. That is not a scandal; it is what vendor-reported numbers look like when nobody in the middle is running the harness for both sides. Which is precisely the role the Intelligence Index plays for Qwen4-Max and cannot play for Bolt.

What the independent board says about Qwen4-Max — and why it does not settle anything
Qwen4-Max's independent record is thin but real. Artificial Analysis scores it 43 on Intelligence Index v5.0.1 and ranks it 51st of 96 models on that version of the board. It costs $3.27 of compute per index task, holds up by generating 127 million output tokens across the run against a median of 65.1 million, and takes 31.3 minutes per task against a median of 12.4. Output speed is 56 tokens per second; time to first answer token is 146 seconds. The model's context window on the board reads 984k, which is the thinking-mode input ceiling of 983,616 rounded, not a different model.

Read those rows as a deployment profile rather than as a verdict. A model that thinks for half an hour per task and emits twice the median token count is not slow because it is badly served; it is slow because it is doing more work per task and billing you for the work. On a per-task basis, $3.27 is the honest cost of a Qwen4-Max answer, and it is the number to carry into any comparison with a model you host yourself, where the marginal cost of a task is electricity and queue time.
Now try to put Bolt's 24GB next to that. You have a vendor table saying it beats a 35B open-weight Qwen on agentic browsing, and you have an independent board saying the hosted Qwen it is actually being compared against scores 43 and costs $3.27 a task. Those two facts do not touch. No one has run Nemotron 3.5 Bolt 74B through the Intelligence Index, and until someone does, any chart placing the two models on a capability axis is drawing a line between a measured point and an assumed one.
Where the comparison is decidable: memory, latency and the bill
Three rows can be compared honestly, because both sides publish them and neither requires a third party.
• Memory — Nemotron 3.5 Bolt 74B-A9B 24GB needs 24 gigabytes of local memory at 2.0625-bit expert quantization; Qwen4-Max needs none, because you never hold the weights. The trade is that a hosted model's memory ceiling becomes a context ceiling: 262,144 tokens for the NVIDIA checkpoint, 1,000,000 for Qwen4-Max, with a 983,616-token input limit in thinking mode and 262,144 shared between reasoning and answer.
• Speed — NVIDIA claims at least 80 tokens per second decoding and reports 84 with the drafter; Artificial Analysis measures Qwen4-Max at 56 tokens per second, with 146 seconds to the first answer token. The local model is faster per token and starts answering sooner. The hosted model may still finish the task faster, because it is doing a different amount of thinking.
• Price — a Qwen4-Max task costs $3.27 on the independent harness, or $1.50 and $4.50 per million tokens at the meter. Nemotron 3.5 Bolt 74B has no published metered price at all, because there is no provider selling it. Its cost is whatever the GPU costs you: a 24GB card that is already in a workstation, amortized over however many tasks a month you run through it.
That last row is where the decision actually lives, and it is why the "which is smarter" framing keeps failing. If your workload is a million short agent turns a month, the local checkpoint wins on unit economics by a margin no benchmark row can overcome, and the 262k context is a constraint you design around. If your workload is a few thousand long-horizon tasks with video in the input and a 900,000-token document attached, no 24GB checkpoint answers the question at all, and Qwen4-Max's $3.27 per task is the cheapest available answer rather than an expensive one.
Why the honest answer is "both, for different jobs"
The tempting resolution is to declare a winner, and the SERP is full of pages that do exactly that with no shared data. A more useful resolution is that these two models are converging on the same architectural idea from opposite directions. Qwen4-Max is a sparse mixture-of-experts that activates 71 billion of its 2.4 trillion parameters per token and leans on token efficiency — the vendor's own unreproduced claim is that it outperforms Qwen3.8-Max using up to about 50% fewer tokens, with average cost per task falling from $4.43 to $1.67. Nemotron 3.5 Bolt 74B is a sparse mixture-of-experts that activates 9 billion of its 74 billion and leans on compression, buying a 24GB footprint with 2-bit expert weights and a small speculative drafter. Both are answers to the same question: how little of the model can you touch per token before quality collapses?
That framing has a practical consequence. The teams that will get the most out of either are the ones measuring tokens per finished task rather than tokens per second. On the local side that means watching acceptance length on the drafter, because a speculative decoder that accepts 4.68 tokens per step under greedy decoding and 4.25 under sampling is where your throughput comes from, and sampling settings change it. On the hosted side it means logging combined thinking-plus-answer totals against the 262,144-token ceiling, because reasoning effort defaults to high and an OpenAI-shaped client asking for medium silently receives high. Neither measurement requires a benchmark, and both are more informative than one.
What would settle it
The missing number is one harness run. If Nemotron 3.5 Bolt 74B-A9B 24GB were put through Intelligence Index v5.0.1 at its shipped quantization, the two models would become comparable on the only axis that both sides can be measured on, and every aggregator page that currently says "no shared benchmark results yet" would have something to print. NVIDIA has published the checkpoint, the runtime is vLLM, and the tooling exists. Until that happens, the useful move is to stop treating this as a matchup and start treating it as a build-versus-buy decision with a documented cost on one side and a documented footprint on the other.
On our own side of the ledger, being precise about availability: OrcaRouter routes neither of these models. Nemotron 3.5 Bolt 74B is a download you serve yourself, and Qwen4-Max is rented from Alibaba or through partner platforms during its current free window. The Qwen family member in our catalogue is Qwen3.8-Max, at $2.00 in and $6.00 out per million tokens with a 1M-token context, which is the natural fallback if you want one key across a mixed stack and a price that follows the vendor's own list rather than a marked-up resale of it. For a decision like this one — where the two candidates sit at opposite ends of the build-versus-buy spectrum — the useful thing a router gives you is not a winner. It is the ability to keep the hosted lane open while you find out what the local lane cannot do.
If you are choosing this month, choose on the constraint you cannot move. Context, modality and paying per task are Qwen4-Max's constraints to satisfy. Memory, offline operation and zero marginal cost per request are Nemotron 3.5 Bolt 74B's. The benchmark gap between them is real, but it is not the gap that will decide your architecture — and no amount of side-by-side table-making will close it before someone runs the harness.
Compared in this article1
Detected from this article · Benchmarks: Artificial Analysis · updated daily
