limite-1b-violetto-GGUF

Limite 1B - Violetto (GGUF)

High-quality GGUF quantizations of Paradigma's Limite 1B - Violetto, built for high-throughput mathematical reasoning.

Limite is a 1-billion parameter dense autoregressive transformer, trained from scratch on a mixture of less than 300B highly curated tokens, capable of handling sequences up to 131k tokens long. It is deliberately lightly instruction-tuned and designed to respond in single turns, with extremely high mathematical capability per parameter count — rivaling models tens of times larger on competition-level math (e.g. 74.25% average on BeyondAIME).

See the original model card and the release blog for the full overview, evaluation tables, and prompting guidance.

GGUF files

For a 1B model the quality cliff below Q4 is steep — Q8_0, Q6_K, Q5_K_M and Q4_K_M are the recommended files.

Quantization method

Following the current best practice for GGUF quantization (per-layer importance-aware K-quants in the spirit of Unsloth's Dynamic approach — pure post-training quantization, no QAT):

• Converted losslessly to BF16 GGUF first, consuming every checkpoint tensor.
• Calibrated with a math-domain importance matrix (GSM8K-training/MBPP corpus, chat-template formatted, ~327k tokens) instead of generic Wikipedia text — critical for a math-reasoning model. Coverage verified for all 336 main projection tensors.
• Sensitive custom-architecture tensors (attention/VE gates, residual coefficients, MUDD mixers, rotary tables) preserved in full precision.
• Quantized with the public Unsloth exporter (unsloth_zoo.llama_cpp.quantize_gguf); Q2_K_L uses the Unsloth mixed-precision preset (q2_k base with ffn_down upcasts, Q6_K output, Q4_K embeddings).

Run

Runtime requirement: Limite is a custom architecture (limite). These files need a llama.cpp build with Limite support — stock llama.cpp / Ollama / LM Studio / Unsloth Desktop builds do not implement this architecture yet and will fail with unknown model architecture: 'limite'. Build from a patched llama.cpp (e.g. the community limite.patch against upstream 5836771) with llama-completion, llama-server, llama-quantize targets.

llama-completion \
  -m limite-1b-violetto-Q4_K_M.gguf -c 16384 --no-context-shift \
  -n 256 -p "What is 27 times 34? Show the steps."

For Violetto chat, use llama-server with the embedded Jinja template and user-only messages; preserve its fixed mathematical system prompt.

Sampling settings: temperature=0.6 and top_p=0.95 (from the upstream generation_config.json).

Prompting and intended use

Limite is designed to respond in single turns with radically different response patterns than regular assistants — it can lose the scope of a prompt and reinterpret it as a different, often mathematical, task. Don't expect general instruction-following; see the release blog for examples.

License and attribution

Model weights: Apache-2.0, carried over from Paradigma's release. This is a community conversion, not an endorsed release from Paradigma. Original citation:

@misc{paradigma2026limite,
  title = {{Limite 1B - Violetto}},
  author = {Prignano, Mario and Cirillo, Gabriele and
            Morosini, Alessio and Cerovaz, Luca and
            Bartolocci, Alessandro and Rodolà, Emanuele and
            Starace, Giulio and Pappone, Francesco},
  year = {2026},
  howpublished = {\url{https://paradigma.inc/blog/limite-1b-violetto/}}
}