Arcle V1
Unified omni model.
One model. Seven capabilities. 5.84 billion parameters. Text, images, documents, speech and audio all flow through one nn.Module and one shared representation space — a single forward pass, a single set of weights, a single model file.
Access the ModelFig. 1 — What it is
Not a pipeline.
One model.
Most systems that advertise "multimodal" are orchestration layers — a vision model here, a speech model there, a language model in the middle, glued together by application code.
Arcle V1 instead projects every modality into the same 2,560-dimensional hidden space that the language core reasons in, so the model genuinely sees an image and hears audio in the same way it reads text. It is not a router, not a pipeline, and not a collection of separate models called one after another behind an API.
Fig. 2 — Specifications
Complete
specifications
| Specification | Value |
|---|---|
| Total parameters | 5,840,000,000 (exact) |
| Language core | Hybrid Mamba-2 SSM + attention |
| Core hidden dimension | 2,560 |
| Context window | 2,097,152 tokens (2M) |
| Image generation | 512 × 512, 8-step LCM sampling |
| Image understanding | at 384 px |
| Document / OCR encoder | 256 tokens per page |
| Speech recognition | Encoder, 30-second windows |
| Speech synthesis | 24 kHz |
| Languages | 18+ |
| Precision | bfloat16 |
Fig. 3 — Every feature
Thirteen
capabilities
Text generation & conversation
Full instruction-following chat with reasoning, explanation, summarisation, rewriting and open-ended dialogue.
Mathematical & logical reasoning
Step-by-step working on multi-step arithmetic, algebra and word problems, reinforced through preference optimisation on mathematics-specific data.
Code generation & understanding
Writes, explains and reasons about code, reinforced with dedicated code preference data.
Text-to-image generation
Original 512×512 images from a prompt in 8 sampling steps — fast enough for interactive use. Conditioned by the language core itself.
Image understanding & VQA
Describes images, answers questions about them, and reasons over visual content.
Document reading & OCR
Reads text out of scanned pages, PDFs, forms, charts and screenshots. 95% content-word recall with no repetition artefacts.
Speech-to-text (ASR)
Transcribes spoken audio through the Whisper Medium encoder feeding a window-level connector. Measured at 9.7% word error rate.
Text-to-speech
Generates natural 24 kHz speech through the Voice engine, conditioned by a trained style connector.
Audio understanding
Spoken input feeds the same reasoning core, so it can be reasoned about, answered and acted on in one pass.
Long-context processing
A 2,097,152-token window enabled by the hybrid Mamba-2 backbone — sequence length handled natively, not at attention's quadratic cost.
Multilingual capability
18+ languages, with deliberate Hindi strength.
Deep India knowledge
Explicitly trained to know India — its geography, culture, history, systems and languages — rather than treating it as a footnote.
Single-file deployment
Every capability lives in one nn.Module. One model load — no service mesh, no inter-model latency, no orchestration layer.
Fig. 4 — Evaluation
Benchmark
scores
| Benchmark | Score | Measures |
|---|---|---|
| GSM8K | 77.5% | multi-step mathematical reasoning |
| ARC-Easy | 80.0% | grade-school science reasoning |
| HellaSwag | 67.0% | commonsense sentence completion |
| ARC-Challenge | 48.5% | hard science reasoning |
| MMLU | 43.5% | broad multi-domain knowledge |
GSM8K at 77.5% is well above what a 3B-class parameter budget would normally suggest — Arcle's own capability from the reasoning-SFT stage, not inherited from base weights.
Fig. 5 — Engineering
How it was
built
Modality-grouped batching
Every batch draws from one modality, turning five tiny sub-batches per step into one full-width forward pass.
Length-bucketed sampling
Near-uniform sequence lengths per batch so padding is not wasted compute, while composition still reshuffles every epoch.
Checkpoint rotation & crash recovery
Training survives interruption and resumes on its own weights.
Exact parameter budgeting
LoRA rank and connector widths solved to land the total on 5,840,000,000 precisely.
Trained over 180 hours on one GPU. It generates images from text. It reads documents. It listens and transcribes. It speaks. It reasons through mathematics. It knows India. And it does all of it in one forward pass through one set of weights.
Free for the people who need it most.
Weights are open on Hugging Face.