ModelsArchitectures & capability
Paradigma releases Limite 1B – Violetto, an open-weight 1B math model it says scores 94% on AIME 2026
Rome-based Paradigma has released Limite 1B – Violetto, an open-weight math solver with about 1 billion parameters under Apache 2.0. It is built to run as one of many cheap parallel solvers inside agent systems, not as a general chatbot.
RisingHow it’s measured
YouTube video · 8.9K views · 72 likes
Reach blends public engagement (views, likes, comments, shares, stars) from the platforms where the story or its original source lives: YouTube, Hacker News, GitHub, Substack and similar. A video and the article written from its transcript share one score, taken from the video. Platforms with no data are left out rather than counted as zero, and the age of a story never lowers its score.
AT Pulse research deskPublished · 4 min read · AI-assisted analysisParadigma reports that Violetto scores 94.01% on AIME 2026, 83.62% on HMMT February 2026 and 74.25% on BeyondAIME. These are the company's own figures, and no outside group has re-run them. Some competitor scores in its comparison table come from other model cards or MathArena rather than re-runs under the same settings. That makes its claimed lead over models of 3B parameters and up a loose ranking, not a controlled one. [4] [11] [3]
The design bet is what Paradigma calls 'high-frequency mathematical intelligence'. Many small solvers attack the same problem in parallel, and a selector keeps the promising answers. The model used just under 2 GB of VRAM in a local test, so many copies can share one GPU. Serving goes through a custom vLLM plugin with an OpenAI-compatible API, and tensor and pipeline parallelism are currently limited to 1. [4] [3] [5] [9]
The model's scope is narrow by design: one math problem per turn, with deliberately light instruction tuning. Paradigma warns that it can turn prompts into different, usually mathematical, tasks. A hands-on test confirmed this. Asked for a haiku, it returned a boxed definition, and asked about photosynthesis, it described unrelated biology. It is a poor fit for the general business tasks some commentary suggests using it for. [4] [5] [9]
an open 1B model posts competition-level math scores in under 2 GB of VRAM. Implication: running many solvers at once becomes cheap. The reviewed sources do not show an ensemble of Violetto copies beating a single large model.
Executive brief
Paradigma, a lab in Rome, says its 1-billion-parameter math model, Limite 1B – Violetto, scored 94.01% on AIME 2026. In the company's own comparison table, that beats small models two to nine times its size (Paradigma blog). The model is meant to be one of many cheap solvers running in parallel inside agent systems, not a chatbot. The weights are open under Apache 2.0. No independent group has re-run the benchmark numbers in the reviewed sources. One hands-on test confirmed that the model solves hard problems and runs in under 2 GB of VRAM. The same test found it drifts badly on anything that isn't math (MindStudio).
What changed and event timeline
Violetto released
Paradigma put out open weights, a base model and vLLM serving code under Apache 2.0. Runtimewire reports the lab built it in six weeks (;).
Trade press repeats the headline
Coverage restated the 94.01% AIME 2026 score and the claim that it beats models from 3B to 120B parameters, without testing anything itself ().
First hands-on local test
On an RTX A6000, the model solved a competition problem with sound reasoning. It failed non-math prompts such as a haiku request ().
Creator commentary
A video by an AI avatar of Julian Goldie and the linked Reddit post retell Paradigma's numbers, then turn to a "manager plus helpers" pitch for a paid community (;).
Capabilities and access
- Model:
paradigma-inc/limite-1b-violetto. Dense, about 1B parameters, 131,072-token context (GitHub; Hugging Face). - Scope: solves one math problem per turn. Instruction tuning is deliberately light (GitHub).
- Serving: a custom vLLM plugin with an OpenAI-compatible API. Plain Transformers support is still pending (GitHub; Latenode).
Read the full section
- Model:
paradigma-inc/limite-1b-violetto. Dense, about 1B parameters, 131,072-token context (GitHub; Hugging Face). - Scope: solves one math problem per turn. Instruction tuning is deliberately light (GitHub).
- Serving: a custom vLLM plugin with an OpenAI-compatible API. Tensor and pipeline parallel sizes are currently limited to 1. Plain Transformers support is still pending (GitHub; Latenode).
- Footprint: about a 2.1 GB file (video). It used just under 2 GB of VRAM in testing (MindStudio).
Technical analysis for researchers and developers
- Training: pretrained from scratch on fewer than 300B curated tokens, then synthetic data, SFT and RL. Estimated compute is 1.71×10²¹ FLOPs (Paradigma blog).
- Architecture: draws on community pre-training "speedrun" designs. The model card mentions a hybrid local/global attention layout (Hugging Face).
- Evaluation: a locked
math-extendedsuite pulls immutable Hugging Face snapshots and logs sampling settings, engine details and diagnostics (GitHub).
Read the full section
- Training: pretrained from scratch on fewer than 300B curated tokens, then synthetic data, SFT and RL. Estimated compute is 1.71×10²¹ FLOPs (Paradigma blog).
- Architecture: draws on community pre-training "speedrun" designs. The model card mentions a hybrid local/global attention layout (Hugging Face).
- Evaluation: a locked
math-extendedsuite pulls immutable Hugging Face snapshots and logs sampling settings, engine details and diagnostics (GitHub). Recommended sampling is temperature 0.6, top_p 0.95. The reviewed sources don't state the number of samples per problem. - Design bet: "high-frequency mathematical intelligence." Many small solvers attack one problem in parallel and a selector keeps the promising answers (video).
Claims and evidence
- AIME 2026 94.01%, HMMT Feb 2026 83.62%, BeyondAIME 74.25%, AIME 2025 90.21%. All are Paradigma's own figures (Paradigma blog; Latenode).
- Apex Shortlist 50.80%; Muse Glimmer 30B at 70% on BeyondAIME. Both figures appear only in the video's retelling of Paradigma's materials (video).
- Competitor scores (VibeThinker 3B 93.85%, Qwen3.5 9B 90.42%, and others) partly come from other model cards or MathArena, not from Paradigma re-runs.
Read the full section
- AIME 2026 94.01%, HMMT Feb 2026 83.62%, BeyondAIME 74.25%, AIME 2025 90.21%. All are Paradigma's own figures (Paradigma blog; Latenode).
- Apex Shortlist 50.80%; Muse Glimmer 30B at 70% on BeyondAIME. Both figures appear only in the video's retelling of Paradigma's materials (video).
- Competitor scores (VibeThinker 3B 93.85%, Qwen3.5 9B 90.42%, and others) partly come from other model cards or MathArena, not from Paradigma re-runs. That makes the "beats every model" ranking a loose comparison, not a controlled one (Paradigma blog; video).
- Independent corroboration: none of the benchmark scores has been reproduced. The only independent evidence is the MindStudio test, which is anecdotal.
Context and prior work
- Small math specialists: VibeThinker (1.5B and 3B) and MiniCPM 5 2B were already near 90% on AIME 2026, by the figures Paradigma cites (video). Violetto adds to a trend; it isn't a break from it.
- Funding context: Paradigma was one of ten European teams selected for the "Next Frontier AI" challenge, which offers up to €26.5M (Runtimewire).
Read the full section
- Small math specialists: VibeThinker (1.5B and 3B) and MiniCPM 5 2B were already near 90% on AIME 2026, by the figures Paradigma cites (video). Violetto adds to a trend; it isn't a break from it.
- Funding context: Paradigma was one of ten European teams selected for the "Next Frontier AI" challenge, which offers up to €26.5M (Runtimewire).
Limitations, safety and contested findings
- Paradigma warns that the model can lose the scope of a prompt and turn it into a different, usually mathematical, task (Paradigma blog).
- The compute figures are estimates, and some comparisons leave out the RL stages (MindStudio).
- The video's practical setups (lead sorting, time-zone planning, a writing helper) are general-purpose jobs.
Read the full section
- Paradigma warns that the model can lose the scope of a prompt and turn it into a different, usually mathematical, task (Paradigma blog). MindStudio saw exactly this: asked about photosynthesis, it described unrelated biology, and asked for a haiku, it returned a boxed definition instead of a poem (MindStudio).
- The compute figures are estimates, and some comparisons leave out the RL stages (MindStudio).
- The video's practical setups (lead sorting, time-zone planning, a writing helper) are general-purpose jobs. Paradigma's scope warnings argue against handing them to this model (video).
Business and practitioner implications
- Where it fits: cheap, parallel verification or solving of well-posed math sub-problems inside agent pipelines.
- Where it doesn't: use as a "math helper" behind a router. Inputs must be clean, single-turn math problems, and the outputs need checking.
- Before adopting: re-run the locked eval suite on your own problem mix and compare against VibeThinker and Qwen3.5 under the same settings.
Read the full section
- Where it fits: cheap, parallel verification or solving of well-posed math sub-problems inside agent pipelines. Its small memory footprint allows many concurrent copies per GPU.
- Where it doesn't: use as a "math helper" behind a router. Inputs must be clean, single-turn math problems, and the outputs need checking.
- Before adopting: re-run the locked eval suite on your own problem mix and compare against VibeThinker and Qwen3.5 under the same settings.
- The headline: "a team beats a genius" is the vendor's design thesis. The reviewed sources don't show an ensemble of Violetto copies outperforming a large model.
The source trail.
Sources (11)
Small AI Model Benchmarks Prove a Team Beats a Genius
Feed excerpt only. A transcript of the video linked from this page is supplied (linked video auto captions; https://www.youtube.com/watch?v=Xo1nKCpAZKE); automatic text may contain errors.
reddit.comNEW Limite 1B - Violetto: High Frequency Model! 🤯
Related coverage; assess separately
reddit.comLinked video: Small AI Model Benchmarks Prove a Team Beats a Genius
linked video auto captions
www.youtube.com