Which are the best openrouter free models?

[HelmutsHelmuts is verified member.]

I am planing to play a little with openrouter.. Which are your favourite free models there?

currently I see these 3 from my openrouter dashboard (when filtering the max price to "0" = "free"):

 

[HelmutsHelmuts is verified member.]

my desktop Hermes Agent shows 4 free models:

 

[HelmutsHelmuts is verified member.]

/grok which ones of these 2 free nVidia models perform better and what are their differences?

1) nvidia/nemotron-3-super-120b-a12b: free
2) nvidia/nemotron-3-ultra-550b-a55b: free

 

[DN AI]

Grok Assistant

I need current public specs and comparisons for those two OpenRouter Nemotron models—starting with official/model-card and listing details.
Pulling OpenRouter model pages and the head-to-head comparison for concrete specs.
Fetching Ultra model-card details and any OpenRouter free-tier performance notes.
Grabbing Ultra’s OpenRouter listing text and any benchmark scores for a cleaner side-by-side.
Short answer
nvidia/nemotron-3-ultra-550b-a55b is the stronger model for quality. nvidia/nemotron-3-super-120b-a12b is the lighter, snappier sibling in the same family. On common reasoning/coding/agent work, Ultra usually wins; Super is often better when you care more about free-tier speed and volume.

Main differences

• Size (MoE): Super ≈ 120B total / 12B active. Ultra ≈ 550B total / 55B active. Same idea (sparse MoE), different capacity.
• Role: Super is the efficient multi-agent / high-throughput option. Ultra is NVIDIA’s larger frontier-style Nemotron 3 for harder reasoning, coding, tool use, and long agent runs.
• Architecture: Both are hybrid Mamba + Transformer LatentMoE with multi-token prediction (MTP). Ultra is essentially the scaled-up model in that line, not a totally different design.
• Context: Both are marketed with very large context (OpenRouter/NVIDIA list up to ~1M tokens). Always check the live OpenRouter free endpoint card—served limits can differ by provider/route.
• Reasoning: Both support reasoning-style use; quality ceiling is higher on Ultra.
• OpenRouter free tier feel: Recent free-endpoint stats often show Super faster/cheaper to run in practice (~higher tok/s, lower latency) than Ultra free. Ultra free is still usable but can feel heavier and may hit rate limits sooner under load.
• Release positioning: Super landed first as the efficient open hybrid MoE; Ultra followed as the much larger active-param model.

Which performs better?

• Overall intelligence / hard tasks: Ultra. Independent aggregate indexes put Ultra clearly ahead of Super (roughly high-30s vs mid-20s on one common intelligence index). That matches the ~4–5× active-parameter jump.
• Coding, multi-step agents, tough reasoning, research-style work: Prefer Ultra when the free route is available and responsive.
• Fast iteration, drafts, simpler chat, many parallel light jobs: Prefer Super—still strong for a free MoE, with better free-tier throughput in practice.
• Long-context / multi-doc: Both are built for long agent memory; Ultra has more headroom on hard long-context reasoning, Super is fine for many long-context chores and may feel snappier on free.

Practical pick for OpenRouter free

• Default: Ultra for “best answer” when quality matters.
• Use Super if Ultra is slow, rate-limited, or you are doing lots of quick loops.
• Both free endpoints can change (limits, routing, logging/ToS on NVIDIA free). Don’t put confidential data on free tiers.
• Performance is task-dependent; a quick A/B on your prompts (domain names, brand briefs, negotiation drafts, research summaries) beats any leaderboard.

Bottom line
Same Nemotron 3 hybrid MoE family: Ultra = bigger brain (55B active), Super = efficient workhorse (12B active). For “which performs better,” choose Ultra; for “which is more practical free daily driver,” often Super.