US Open-Weight LLM 2026: Inkling vs Nemotron 3 vs gpt-oss

On July 15, 2026, Thinking Machines shipped Inkling, an open-weight multimodal model, and it immediately took the top spot on the Artificial Analysis Intelligence Index among US-lab open-weight models. That's notable on its own. What makes it a genuine story is the timing: Inkling unseated NVIDIA's Nemotron 3 Ultra just weeks after Nemotron 3 Ultra had taken that same crown from gpt-oss. Three US labs, three leadership changes, inside about six weeks.

If you're a procurement lead at a bank, a CTO at a law firm, or a security engineer evaluating vendors, the benchmark churn is almost a distraction. The question your compliance committee actually asks is narrower: which model was built by a US lab, under what license, and can we fine-tune it on our own data without it leaving our infrastructure. This piece answers that question first, then gets into the capability comparison.

The new US open-weight leaderboard: what changed on July 15, 2026

Before Inkling, the US open-weight conversation was a two-horse race: NVIDIA's Nemotron 3 Ultra (published training recipes, strong audit tooling) versus OpenAI's gpt-oss (smaller, cheaper to run, Apache 2.0 licensed). Inkling adds a third option with a different distribution model entirely: it's natively multimodal, and you can access it either as downloadable weights or through Thinking Machines' Tinker API, which handles the fine-tuning infrastructure for you while still letting you export the resulting weights.

That matters for regulated buyers because it changes the risk calculus. A downloadable-weights-only model means you own every step but you also own every operational headache. A hosted-fine-tuning API means faster time to production but a vendor dependency you have to underwrite. Inkling is the first top-tier US open-weight model to offer both paths natively, which is arguably the more consequential part of the July 15 release, more than the Intelligence Index number itself.

Why country of origin matters for regulated buyers

Raw benchmark score is table stakes. It is not what kills a deal in a bank's vendor risk committee. Three things do:

Procurement policy language. A growing number of financial institutions and government contractors have internal policies (sometimes explicit, sometimes informal) that restrict foreign-origin AI models, particularly from jurisdictions subject to export control scrutiny, from touching material non-public information or client PII.

Export control optics. Even when a Chinese-origin open-weight model (DeepSeek, Qwen, Kimi, GLM) is technically permissible to self-host, boards are wary of the reputational and audit questions that come with defending that choice to regulators or examiners.

Board-level vendor risk. Legal and compliance teams increasingly want a one-line answer to who built this model and under what jurisdiction before capability comparisons even start. That single line can eliminate most of the open-weight field before a technical evaluation begins.

This is the actual reason US open-weight LLM 2026 is a search query at all: it's a filter, not a preference.

Inkling: architecture, benchmarks, and Tinker API vs downloadable weights

Inkling is a mixture-of-experts multimodal model, roughly 180B total parameters with about 24B active per token, a 256K context window, and native image and audio understanding alongside text. It currently tops the US-lab segment of the Artificial Analysis Intelligence Index at 74, ahead of Nemotron 3 Ultra's 71 and gpt-oss's 69.

The Tinker API is the differentiator. It gives you managed fine-tuning infrastructure, so a five-person engineering team can run a LoRA or full fine-tune job without standing up their own training cluster, and you can still export the resulting weights afterward rather than being locked into inference through Thinking Machines forever. For a law firm or bank that wants speed now and full ownership later, that's a meaningfully different posture than a pure weights-download release.

The tradeoff: Inkling is new. There's less third-party audit history, fewer battle-tested fine-tuning recipes in the wild, and its license (a permissive but not fully Apache-style grant with some field-of-use language) needs a careful read from your legal team before you commit production workloads to it.

Nemotron 3 Ultra: published training data, recipes, and eval tooling

Nemotron 3 Ultra remains the strongest choice for teams that need to document exactly what a model was trained on, because NVIDIA publishes the training data composition (the Nemotron-CC corpus lineage), the training recipe, and a compatible eval harness (NeMo Eval) that maps cleanly to internal model risk management documentation. At roughly 340B total parameters with 68B active in its MoE configuration and a 128K context window, it's larger and slower to serve than Inkling or gpt-oss, but it's the model most likely to satisfy a bank's model risk committee on paper, because show me your training data provenance is a question Nemotron 3 Ultra actually has a documented answer to.

gpt-oss: where it fits for US-only stacks

OpenAI's gpt-oss (117B total parameters, about 5.1B active, 128K context, Apache 2.0 license) is the smallest and cheapest of the three to run at scale. It's not going to top the Intelligence Index against Inkling or Nemotron 3 Ultra, but for teams that need a genuinely permissive license, low inference cost, and the shortest path to a self-hosted endpoint, it's still the pragmatic default. If your use case is internal support tooling, code review assistance, or document classification rather than frontier reasoning, gpt-oss frequently wins on cost per token even after you account for the capability gap.

Head-to-head table

When a Chinese-origin model is still the better technical call

Being honest about tradeoffs matters more than being patriotic about them. DeepSeek V3.1, Qwen 3.6, Kimi K2.6, and GLM all currently outperform at least one of these three US models on specific benchmarks, particularly coding and long-context retrieval, and they're often cheaper to serve per token. If your workload has no data sovereignty or export-control constraint (an internal engineering copilot with no client data, for example), ruling out a Chinese-origin model purely on principle is leaving capability and cost savings on the table.

The honest framing: country of origin is a hard filter for regulated data, not a universal ranking criterion. A security research team building an internal tool with synthetic data has a different calculus than a bank fine-tuning on client transaction histories.

Fine-tuning any of these three on your own business data

Whichever model you land on, the harder problem isn't picking Inkling versus Nemotron 3 Ultra versus gpt-oss. It's standing up fine-tuning infrastructure, tracking training and inference spend across whichever GPU provider you use, and getting a production endpoint that your existing tools can actually call.

This is the gap InfoPlatform.ai is built to close. You upload your data, pick Nemotron 3, gpt-oss, or Inkling (via the Tinker API), and pick a training provider (serverless GPUs, dedicated GPUs, or Tinker directly), and you get back an OpenAI-compatible production endpoint that drops into OpenCode, Cursor, or LangChain with a one-line base_url swap. You keep the weights, the training data, and the inference layer. You can also configure delete-training-data-after-training for workloads where retention itself is the compliance risk, which matters a lot if you're fine-tuning on client PII or privileged legal documents.

If you're already running multiple providers for training and inference, keeping a single view of spend across them (rather than reconciling five separate GPU bills) is worth doing before you scale past a pilot. A tool like AICosts.ai exists specifically for that: one dashboard across 50+ providers with budget alerts, which becomes useful the moment you're running Nemotron 3 fine-tunes on one provider and gpt-oss inference on another.

A decision framework: compliance requirement vs. raw capability vs. cost

Use this order, not the reverse:

Filter by jurisdiction first. If your data has regulatory sensitivity (client financials, PII, privileged legal material), narrow to US-lab models before comparing anything else.

Match modality to workload. If you need image or document understanding alongside text, Inkling is currently the only one of the three built for it natively.

Match audit needs to documentation. If your model risk management process demands training data provenance, Nemotron 3 Ultra's published recipes make your compliance write-up materially easier.

Optimize cost last, not first. gpt-oss's smaller active parameter count means meaningfully cheaper inference at scale, which matters once you're past pilot and into production volume.

Confirm you can actually own the output. Downloadable weights plus a delete-after-training option beats a hosted-only endpoint for any team that needs to prove data never left their control.

FAQ

Is Inkling actually made in the US?

Yes. Inkling is built by Thinking Machines, a US AI lab, and released as an open-weight model on July 15, 2026, with distribution through both downloadable weights and the Tinker API. It currently leads the US-lab segment of the Artificial Analysis Intelligence Index.

Is Nemotron 3 Ultra open-weight?

Yes. NVIDIA releases Nemotron 3 Ultra's weights under the NVIDIA Open Model License, along with training data composition details and a compatible evaluation harness, which is why it remains a strong choice for regulated teams that need documented training provenance for model risk management.

Is gpt-oss safe for a bank or law firm to self-host?

Gpt-oss is released by OpenAI under an Apache 2.0 license, which is one of the most permissive open-weight licenses available, and it can be fully self-hosted with no data leaving your infrastructure. It's a reasonable choice for cost-sensitive, US-origin workloads that don't require frontier-level reasoning or multimodal input.

Why not just use DeepSeek, Qwen, or Kimi if they benchmark higher on some tasks?

For workloads without regulatory data sensitivity, a Chinese-origin model can be the better technical and cost choice. For workloads touching client PII, financial data, or privileged legal material, procurement policy and export-control optics at most regulated institutions rule these out regardless of benchmark performance, which is why country of origin needs to be the first filter, not the last one.

Can I fine-tune Inkling, Nemotron 3, or gpt-oss without an ML team?

Yes. InfoPlatform.ai lets you upload your data, select any of these three models, choose a training provider (serverless GPUs, dedicated GPUs, or Tinker), and get back a production OpenAI-compatible endpoint, while you retain full ownership of your data, weights, and inference. Get started to fine-tune your first model on your own business data.