Inkling vs Nemotron 3 Ultra: Which Open-Weight Model Should You Deploy?
This Inkling vs Nemotron comparison pits two open-weight models against each other: Inkling, the debut release from Thinking Machines Lab (the startup led by former OpenAI CTO Mira Murati), and Nemotron 3 Ultra, NVIDIA’s flagship open model. Both ship downloadable weights, both target teams that want to self-host and fine-tune rather than rent a closed API, and both play in the same open-weight tier. The interesting part: across the head-to-head figures we have, Inkling vs Nemotron 3 Ultra is the one open matchup where Inkling leads on every benchmark row in our data. Below we lay out the numbers honestly, then cover licensing, VRAM, cost, and where NVIDIA’s stack still gives Nemotron a real edge.
A note for builders: there are no audited head-to-head benchmarks here, so this compares models and access, not scores. OrcaRouter routes API-available models behind a single OpenAI-compatible endpoint, so you can trial and compare Inkling and Nemotron 3 Ultra without wiring up multiple SDKs.
TL;DR verdict: Pick Inkling if you want the stronger raw scores in our data, a permissive Apache 2.0 license, a 1M-token context window, and multimodal (text + image + audio) input. Pick Nemotron 3 Ultra if you are standardized on NVIDIA’s enterprise and hardware stack (NIM microservices, NeMo, certified DGX/Blackwell deployments) and want a model tuned to slot into that ecosystem. Both are open-weight and self-hostable.
Key takeaways
Both are open-weight, downloadable, and self-hostable — this is an open-vs-open matchup, not open-vs-closed.
Inkling leads every benchmark row in our head-to-head data (MarkTechPost set), from HLE and AIME 2026 to SWE-bench Verified, Terminal Bench 2.1, and FORTRESS.
Inkling also leads the independent index: Artificial Analysis Intelligence Index 41 vs 38 for Nemotron 3 Ultra.
License difference: Inkling is Apache 2.0; Nemotron 3 Ultra ships under NVIDIA’s open model license — check NVIDIA’s terms for the specifics before commercial deployment.
Nemotron’s edge is positioning: NVIDIA’s enterprise/hardware stack integration, not benchmark wins in our data.
Caveat: Inkling’s benchmarks are vendor self-reported at launch; competitor figures come from third parties and are not independently audited.
Benchmarks here are vendor self-reported at launch (Effort 0.99) for Inkling, and third-party figures are from Artificial Analysis and MarkTechPost; none are independently audited, and competitor numbers may differ from NVIDIA’s own reported figures. Inkling’s own specs are from Thinking Machines’ model card.
Quick-glance comparison
Maker. Inkling: Thinking Machines Lab; Nemotron 3 Ultra: NVIDIA
License. Inkling: Apache 2.0 (royalty-free self-host); Nemotron 3 Ultra: NVIDIA open model license (check NVIDIA terms)
Weights. Inkling: Open (Hugging Face); Nemotron 3 Ultra: Open
Params. Inkling: 975B total / 41B active (MoE); Nemotron 3 Ultra: Not in our data
Context. Inkling: Up to 1M tokens (256K on hosted APIs); Nemotron 3 Ultra: Not in our data
Modalities. Inkling: Text + image + audio in, text out; Nemotron 3 Ultra: Not in our data
Self-host / fine-tune. Inkling: Yes / Yes (Tinker); Nemotron 3 Ultra: Yes / Yes
Hosted price. Inkling: ~$1.87 in / ~$4.68 out per 1M (AA); Nemotron 3 Ultra: Not in our data
Cells marked “Not in our data” are omitted rather than guessed — see the disclosure above.
Winner by category
Reasoning / Knowledge. Winner: Inkling; Notes: HLE 29.7% vs 26.6% (MarkTechPost)
Math. Winner: Inkling; Notes: AIME 2026 97.1% vs 94.2%
Coding. Winner: Inkling; Notes: SWE-bench Verified 77.6% vs 70.7%
Agentic (terminal). Winner: Inkling; Notes: Terminal Bench 2.1 63.8 vs 56.4
Safety (adversarial). Winner: Inkling (narrow); Notes: FORTRESS 78.0% vs 77.6%
Overall intelligence. Winner: Inkling; Notes: AA Intelligence Index 41 vs 38
Multimodal / Audio. Winner: Inkling; Notes: Text+image+audio in; Nemotron modalities not in our data
Enterprise/hardware fit. Winner: Nemotron 3 Ultra; Notes: Native NVIDIA stack integration
Cost (self-host). Winner: Tie; Notes: Both royalty-free to self-host (per each license)
Head-to-head benchmarks
The following figures come from a single consistent set reported by MarkTechPost, plus one independent index from Artificial Analysis. Bold marks the leader.
HLE (no tools). Inkling: 29.7%; Nemotron 3 Ultra: 26.6%; Source: MarkTechPost
AIME 2026 (math). Inkling: 97.1%; Nemotron 3 Ultra: 94.2%; Source: MarkTechPost
SWE-bench Verified (coding). Inkling: 77.6%; Nemotron 3 Ultra: 70.7%; Source: MarkTechPost
Terminal Bench 2.1 (agentic). Inkling: 63.8; Nemotron 3 Ultra: 56.4; Source: MarkTechPost
FORTRESS (adversarial). Inkling: 78.0%; Nemotron 3 Ultra: 77.6%; Source: MarkTechPost
AA Intelligence Index. Inkling: 41; Nemotron 3 Ultra: 38; Source: Artificial Analysis
This is a clean sweep for Inkling in the data we have. It is worth stating plainly: among the open rivals Inkling was benchmarked against, Nemotron 3 Ultra is the one it beats across the board. Against other open models like GLM 5.2, Kimi K2.6, and DeepSeek V4 Pro, Inkling trades wins and losses — but here it leads every row.
Keep the caveats in view, though. These are launch-day, self-reported numbers on Inkling’s side, and the competitor scores were compiled by third parties rather than independently audited. The margins on FORTRESS (78.0% vs 77.6%) are narrow enough that a re-run under different harness conditions could flip them. Treat the direction as more reliable than the decimals.
Where Nemotron 3 Ultra wins
Nemotron 3 Ultra’s advantage is not on the scoreboard in our data — it is positioning. Nemotron is NVIDIA’s own model family, and that carries real weight for enterprises already committed to NVIDIA’s stack:
Hardware and software co-design. Nemotron models are built to run cleanly on NVIDIA’s hardware and are surfaced through NVIDIA’s enterprise tooling (NIM inference microservices, the NeMo framework, and certified DGX/Blackwell reference deployments). If your platform team already runs on that stack, Nemotron slots in with the least friction.
Enterprise support and packaging. A model backed by NVIDIA’s commercial machinery is an easier procurement and support story for large organizations than a first release from a young startup.
Ecosystem gravity. For teams standardizing on one vendor for GPUs, drivers, inference runtime, and model, Nemotron reduces the number of moving parts.
None of that shows up in a benchmark table, but it is often the deciding factor in enterprise deployments.
Where Inkling wins
Every benchmark in our data. HLE, AIME 2026, SWE-bench Verified, Terminal Bench 2.1, and FORTRESS all favor Inkling, as does the independent AA Intelligence Index (41 vs 38).
More permissive license. Apache 2.0 is about as unrestrictive as open licensing gets. Nemotron’s NVIDIA open model license may carry conditions worth reviewing (see below).
Multimodal input. Inkling accepts text, images, and audio in (text out). Nemotron’s modality support is not in our data.
Huge context window. Inkling’s weights support up to 1M tokens (256K on hosted APIs).
Controllable thinking effort. A reasoning-effort dial lets you trade cost for depth per request.
Pricing & cost / TCO
Because both models are open-weight, the headline cost question is the same for each: self-hosting is royalty-free (subject to each model’s license terms). You pay for the GPUs and the ops, not for the weights.
For Inkling, if you prefer managed hosting, third-party providers price it (per Artificial Analysis) at roughly $1.87 / 1M input tokens and $4.68 / 1M output tokens at 64K context (cache around $0.374 / 1M), rising to about $3.74 / $9.36 at 256K context. Fine-tuning runs through the Tinker platform (64K and 256K context options), with a 50% limited-time launch discount, and there is a free Playground to try it. Inkling is also notably token-efficient (~25K output tokens/task), which lowers real-world output-token spend.
For Nemotron 3 Ultra, we do not have hosted per-token pricing in our data, so we won’t quote a number. Qualitatively: if you run it inside an existing NVIDIA enterprise agreement, the model cost may fold into a broader stack deal, which can change the TCO math independently of any per-token rate.
Licensing & deployment
Licensing. Inkling is released under Apache 2.0 — commercial use is permitted, self-hosting is royalty-free, and the terms are simple and well understood. Nemotron 3 Ultra ships under NVIDIA’s open model license. We are not going to guess at its specific clauses; the responsible move is to read NVIDIA’s terms directly before committing to commercial deployment, since open-model licenses can include use restrictions, attribution requirements, or acceptable-use conditions that Apache 2.0 does not. The practical takeaway: Inkling’s license is the more permissive and predictable of the two.
How to run Inkling. The weights are on Hugging Face (BF16 + an NVFP4 checkpoint for NVIDIA Blackwell). VRAM tiers:
BF16: ~2TB (roughly 8×B300 or 16×H200)
NVFP4: ~600GB (roughly 4×B300 or 8×H200)
Constrained setups: Unsloth 1-bit GGUF quantizations
Supported runtimes include SGLang, vLLM, TokenSpeed, Unsloth, and Hugging Face transformers, and hosted access is available through Together AI, Fireworks, Modal, Databricks, and Baseten. A minimal vLLM quickstart looks like:
vllm serve thinkingmachines/Inkling --tensor-parallel-size 8
How to run Nemotron 3 Ultra. Nemotron 3 Ultra is likewise open-weight and self-hostable, and it is designed to run through NVIDIA’s own deployment path (NIM microservices and the NeMo framework on NVIDIA hardware). We don’t have its exact VRAM footprint or a per-token price in our data, so check NVIDIA’s model page for the checkpoint sizes and supported runtimes.
Which should you choose?
Choose Inkling if you want the stronger measured performance in our data, the most permissive license (Apache 2.0), multimodal input, a 1M-token context window, and a token-efficient model you can fine-tune on Tinker. It is the better pick for cost-sensitive teams and anyone who wants maximum flexibility over how they deploy.
Choose Nemotron 3 Ultra if your organization is already standardized on NVIDIA’s enterprise and hardware stack and values that tight integration, packaging, and support over the benchmark gap. The scoreboard favors Inkling; the ecosystem may favor Nemotron for you.
Not sure? Both are free to self-host, so the low-risk move is to prototype Inkling (via the free Playground or a hosted provider) and Nemotron (via NVIDIA’s deployment path) on your own representative tasks. Benchmarks point one way; your workload is the real judge.
For a deeper look at Inkling itself, see our full Inkling AI model review and the explainer What is Inkling AI?. For other open-weight matchups, compare Inkling vs Kimi K2.6 and Inkling vs GLM 5.2, where the results are closer than they are here.
FAQ
Is Inkling better than Nemotron 3 Ultra? On the data we have, yes. Inkling leads every head-to-head benchmark row (MarkTechPost set) and the independent Artificial Analysis Intelligence Index (41 vs 38). That said, these figures are self-reported or third-party and not independently audited, and “better” also depends on how well each model fits your existing stack.
Which is better for coding? Inkling, per the numbers: SWE-bench Verified 77.6% vs 70.7% and Terminal Bench 2.1 63.8 vs 56.4 (both MarkTechPost). As always, validate on your own codebase before deciding.
Which is cheaper? Both are royalty-free to self-host, so the honest answer is “it depends on your infrastructure.” Inkling has a published hosted price (~$1.87/$4.68 per 1M input/output tokens via AA) and is token-efficient; we don’t have Nemotron’s hosted pricing in our data, and its cost may fold into a broader NVIDIA agreement.
Is Nemotron 3 Ultra open source? It is open-weight — the weights are downloadable — but it ships under NVIDIA’s open model license, not a standard OSI-approved open-source license. “Open weight” is not the same as “open source.” Check NVIDIA’s terms for the specifics. Inkling, by contrast, uses Apache 2.0.
Can I self-host Nemotron 3 Ultra? Yes. It is open-weight and self-hostable, designed to run through NVIDIA’s deployment tooling (NIM/NeMo) on NVIDIA hardware. Review the license before commercial use.
Can I fine-tune Inkling? Yes. Inkling is built for customization: fine-tune via the Tinker platform (64K/256K context options, with a launch discount) or self-host the Apache 2.0 weights and fine-tune on your own infrastructure.
Conclusion
Among the open rivals we have data for, Nemotron 3 Ultra is the one Inkling beats cleanly — leading every benchmark row and the independent intelligence index, with a more permissive Apache 2.0 license and multimodal, long-context support on top. Nemotron 3 Ultra’s real advantage is not the scoreboard but its native fit inside NVIDIA’s enterprise and hardware ecosystem, which can matter more than a few benchmark points for teams already committed to that stack. Keep the caveats in mind — none of these numbers are independently audited — but if you are choosing on measured capability and licensing freedom, Inkling is the stronger pick here.
