What is Limite 1B - Violetto?
Limite 1B - Violetto is the first model from Paradigma, a research start-up in Rome whose founders include Sapienza professor Emanuele Rodolà and former OpenAI benchmark builder Giulio Starace. It is a 1-billion-parameter dense transformer trained from scratch on under 300 billion curated tokens, with an architecture borrowed from community pre-training speedrun work (a mix of local and global attention layers), then post-trained with synthetic data, supervised fine-tuning and reinforcement learning. It handles sequences up to 131,072 tokens.
It does one thing: solve competition mathematics in a single turn. Paradigma deliberately kept instruction tuning light, so it is not a chat assistant and will often turn an ordinary question into a maths problem. Weights and the vLLM serving plugin are under Apache 2.0; a technical report is promised but not yet published.
Is Limite 1B - Violetto worth it?
On Paradigma's own table, Limite does competition maths at a level no 1B model has reached. It scores 94.01% on AIME 2026, 83.62% on HMMT February 2026 and 74.25% on BeyondAIME, against 93.85%, 78.98% and 72.00% for VibeThinker-3B, three times its size, and 93.44%, 80.87% and 70.00% for the 30B MUSE-Glimmer. MiniCPM5-2B, the strongest small generalist in the table, gets 90.21% on AIME 2026 and trails Limite on all seven benchmarks; VibeThinker-3B is ahead only on AIME 2025 and ArXivMath. GLM 5.2 and Gemini 3 Pro still lead on the hardest set, APEX Shortlist, at 68.09% and 66.49% against Limite's 50.80%.
Two cautions apply. Every Limite figure is self-reported, the scores for several rivals were copied from model cards or MathArena rather than rerun, and the footnote warns that evaluation settings differ. The technical report that would explain the training data is still pending, which matters because contamination is the usual worry with a model this small scoring this high on 2026 contests.
The narrowness is a design choice, and Paradigma is candid about it. Its own examples show the model answering "What is photosynthesis?" with a description of cell division and treating a quoted injection string as a sequence to differentiate. The card asks for one question per conversation, the canonical system prompt and no history. That rules it out as a general assistant or a component that sees untrusted text, but suits a pipeline that only ever sends it maths.
In deployment, it is cheap to run but fussy. The recommended path is Paradigma's vLLM 0.26 plugin on CUDA 13 with an NVIDIA GPU; Transformers works with trust_remote_code but was validated only on an H100, and neither path supports sharding or fine-tuning officially. The card lists no GGUF or Apple builds.
Where it earns its place
An ML researcher studying small-model reasoning
Generate thousands of sampled solutions per problem to study best-of-n and verification
Serve it with the vLLM plugin on one GPU at temperature 0.6 and sample widely; at 1B parameters, throughput per card is far higher than with the 27B to 30B models it rivals. Check answers with a verifier, since scores are self-reported.
An online maths-olympiad coaching platform
Draft worked solutions for practice problems before a coach reviews them
Send one problem per request with the canonical system prompt and no history. Keep it behind a review step and away from free-text student chat, where it drifts into unrelated maths.
Before you commit
- Setup
- git clone paradigma-inc/limite-violetto, uv sync --locked, then VLLM_PLUGINS=limite uv run vllm serve; needs Linux x86-64, an NVIDIA GPU and a CUDA 13-compatible driver. Transformers works with trust_remote_code.
- Integrations
- OpenAI-compatible serving through vLLM; Transformers with SDPA or FlashAttention 2 (not FlashAttention 2 with a static cache). No GGUF, Ollama or MLX builds listed on the card.
- Data and privacy
- Self-hosted only, so problems and answers stay on your hardware. Training data is described only as under 300 billion curated tokens; details await the technical report.
- Cost at scale
- Weights free; a 1B BF16 checkpoint is about 2 GB, so one GPU serves it, with tensor and pipeline parallelism kept at 1 per the card. No hosted price exists.
- Limits
- Single-turn use only; loses the scope of non-maths prompts; fine-tuning, LoRA and multi-GPU sharding are not supported officially.
What’s good
- 94.01% on AIME 2026 at 1B parameters, per Paradigma
- Ahead of 3B to 30B models on most of its seven maths benchmarks
- Apache 2.0 weights and serving plugin
- 131K-token context for long solutions
What’s not
- Self-reported scores; technical report not yet out
- Reinterprets non-maths questions as maths
- Single-turn only; poor fit for untrusted input
- Officially NVIDIA and CUDA 13 only; no fine-tuning support
Best for
Researchers and education-tech teams that need many competition-maths solutions cheaply and can verify answers automatically.
Limite 1B - Violetto alternatives
Limite 1B - Violetto: common questions
- Is Limite 1B - Violetto worth it?
- Remarkable maths per parameter, useless outside maths, and so far graded only by the people who made it. We rate it 3.6 out of 5.
- How much does Limite 1B - Violetto cost?
- Free open weights under Apache 2.0. No hosted API; you pay only for your own GPU.
- What are the downsides of Limite 1B - Violetto?
- Self-reported scores; technical report not yet out; reinterprets non-maths questions as maths; single-turn only; poor fit for untrusted input; officially NVIDIA and CUDA 13 only; no fine-tuning support.
- Who is Limite 1B - Violetto best for?
- Researchers and education-tech teams that need many competition-maths solutions cheaply and can verify answers automatically.
- How much does Limite 1B - Violetto cost for a team or in production?
- Weights free; a 1B BF16 checkpoint is about 2 GB, so one GPU serves it, with tensor and pipeline parallelism kept at 1 per the card. No hosted price exists.
- Is Limite 1B - Violetto safe for company data?
- Self-hosted only, so problems and answers stay on your hardware. Training data is described only as under 300 billion curated tokens; details await the technical report.
- What are the best Limite 1B - Violetto alternatives?
- The closest alternatives we have reviewed are MiniCPM5-2B, Qwen3.8-27B and SmolLM3-3B.
Answers checked Nov 2, 2026.