Skip to main content

Limite 1B - Violetto (Paradigma Inc, local)

Out of scope, added deliberately

Every other model in this registry targets fashion/media (try-on, image, video, understanding, speech). Limite 1B - Violetto does not -- it is a math-reasoning-only text LLM with no image/video/audio capability. It was added on request as a reference point for a possible future "generic model gateway" direction, not because it fits the current OpenTryOn roadmap. If you're looking for a fashion-domain text/VLM model, see Kimi-VL, Qwen3.8, DeepSeek-VL2, or Ternary Bonsai 2 27B instead.

Limite 1B - Violetto is Paradigma Inc's first model: a dense 1B-parameter autoregressive transformer trained from scratch on under 300B curated tokens specifically for high-throughput mathematical reasoning. No first-party hosted API exists -- Paradigma publishes weights only -- so local is the only path.

Overview​

FeatureValue
Modelparadigma-inc/limite-1b-violetto
Parameters1B, dense
LicenseApache-2.0
ContextUp to 131K tokens
Disk / VRAM~2.1GB -- the lightest local model in this registry
CapabilityText-only math reasoning. No image, video, or audio input.
Usage patternSingle-turn only (system prompt + current user message) -- the model card warns multi-turn history may get reinterpreted as a different math problem

Installation​

pip install opentryon[local]   # torch, transformers, etc.
note

The model card recommends Python 3.12 / torch 2.11 (CUDA 13) / transformers 5.6.2 and attn_implementation="sdpa" (FlashAttention2 + StaticCache is explicitly called out upstream as numerically incorrect for this model -- this adapter always uses sdpa). This repo's shared opentryon[local] pin is older (transformers==4.42.4) -- upgrade in a separate env if loading fails with a trust_remote_code error.

Quick Start​

from tryon.models.limite import LimiteAdapter

adapter = LimiteAdapter()  # downloads paradigma-inc/limite-1b-violetto on first use

result = adapter.understand(prompt="If x + 3 = 8, what is x?")
print(result["text"])

Using the opentryon CLI​

opentryon understand --model limite-1b-violetto \
  --prompt "A train travels 60 miles in 1.5 hours. What is its average speed?"

The CLI checks that torch is installed before attempting to load the model (pip install opentryon[local] hint if missing).

API Reference​

LimiteAdapter​

class LimiteAdapter:
    def __init__(
        self,
        model_id: Optional[str] = None,   # LIMITE_MODEL_ID or paradigma-inc/limite-1b-violetto
        device: Optional[str] = None,     # passed as device_map, defaults to "auto"
        torch_dtype: str = "auto",
        trust_remote_code: bool = True,   # required for Limite's custom modeling code
    )

understand​

def understand(
    self,
    prompt: str,
    system_prompt: Optional[str] = None,
    max_new_tokens: int = 512,
    temperature: Optional[float] = None,  # default 0.6 from generation_config.json if omitted
    top_p: Optional[float] = None,        # default 0.95 from generation_config.json if omitted
) -> Dict[str, Any]  # {"text": ..., "model": ...}

Only prompt (and optionally system_prompt) are sent per call -- no chat history -- matching the model card's recommended usage.

MCP​

Same registry model appears as an MCP tool (no extra wiring):

  • understand_limite_1b_violetto -- local GPU, no cloud key

Not wired into the planner's capability defaults or a new intent (it stays named-model-only: --model limite-1b-violetto or "use limite-1b-violetto" in chat) since it doesn't fit any fashion/media intent as a default.

References​