ArcleIntelligence — AI for everyone

Absolute
autonomy.

We build free, private, on-device AI for the people who cannot afford subscriptions. Our first model, Arcle V1, is a unified omni model. Our next architecture, GYRE, aims to change the economics of the whole field.

Access Arcle V1Explore Products
Fig. 0 — The Vision
2026
Model 01Shipped

Arcle V1

Unified omni model

Our first model. Text, images, documents, speech and audio through one 5.84-billion-parameter forward pass — one set of weights, one model file. 77.5% on GSM8K and a 2-million-token context.

5.84BParameters
2M tokensContext
77.5%GSM8K
Read the full story→
Next
Model 02In Research

GYRE Architecture

The industry game-changer

A looped, adaptively-halting MoE-SSM with exact retrieval and a contracted compute budget. Linear-time mixing, no KV cache, exact non-decaying retrieval — attacking the three costs that dominate long-context AI. Coming soon.

3.03BParameters
100% @ 12MRetrieval
75.5× @ 1MMixing
Read the full story→
5.84B Parameters✦
2M Token Context✦
Unified Omni Model✦
One Forward Pass✦
Hybrid Mamba-2 + Attention✦
GSM8K 77.5%✦
Text · Image · Audio · Document✦
18+ Languages✦
Deep India Knowledge✦
Single Model File✦
5.84B Parameters✦
2M Token Context✦
Unified Omni Model✦
One Forward Pass✦
Hybrid Mamba-2 + Attention✦
GSM8K 77.5%✦
Text · Image · Audio · Document✦
18+ Languages✦
Deep India Knowledge✦
Single Model File✦

Fig. 1 — Evaluation

Arcle V1
Benchmarks

Measured against published results for open models in the same parameter class. Reported in full — including where Arcle V1 trails.
77.5
38.4
75.6
80.6
88.6
Arcle V1
5.84B
Gemma 3
4B
Llama 3.2
3B-It
Qwen2.5
3B-It
Phi-4-mini
3.8B
GSM8K
multi-step mathematical reasoning
80.0
82.4
n/a
n/a
n/a
Arcle V1
5.84B
Gemma 3
4B
Llama 3.2
3B-It
Qwen2.5
3B-It
Phi-4-mini
3.8B
ARC-Easy
grade-school science reasoning
67.0
77.2
77.2
74.6
69.1
Arcle V1
5.84B
Gemma 3
4B
Llama 3.2
3B-It
Qwen2.5
3B-It
Phi-4-mini
3.8B
HellaSwag
commonsense sentence completion
48.5
56.2
76.1
82.6
83.7
Arcle V1
5.84B
Gemma 3
4B
Llama 3.2
3B-It
Qwen2.5
3B-It
Phi-4-mini
3.8B
ARC-Challenge
hard science reasoning
43.5
59.6
61.8
65.0
67.3
Arcle V1
5.84B
Gemma 3
4B
Llama 3.2
3B-It
Qwen2.5
3B-It
Phi-4-mini
3.8B
MMLU
broad multi-domain knowledge
Standout result
77.5%
GSM8K

Mathematical reasoning well above what a 3B-class parameter budget would normally suggest — and it beats Gemma 3 4B outright. It comes from the reasoning-SFT stage and the adapter: Arcle's own capability, not inherited from base weights.

Capability measurements — Arcle V1

95.0%
OCR — in-distribution
content-word recall
95.4%
OCR — adapter active
content-word recall
90.3%
Speech recognition
9.7% word error rate
89.6%
Audio pathway
understanding
57.0%
Image understanding
caption overlap
2.09M
Context window
tokens

Methodology — Arcle V1 figures are measured through Arcle's own evaluation harness at n = 400 samples per benchmark, after the reasoning-SFT stage. Peer figures are published results: Gemma 3 4B from the Gemma 3 technical report; Llama 3.2 3B-Instruct, Qwen2.5 3B-Instruct and Phi-4-mini from the Phi-4-mini-instruct model card. Because harnesses, sample counts and shot counts differ between sources, these comparisons are indicative rather than like-for-like. Blank bars mark benchmarks with no published figure from that source.

Fig. 2 — Capabilities

One Model.
Seven capabilities.

Every modality projects into the same 2,560-dimensional space. Not a router, not a pipeline — a single forward pass through one set of weights.
01Instruction-tuned

Text & Reasoning

  • ↳Full instruction-following conversation
  • ↳Explanation, summarisation, rewriting
  • ↳Step-by-step arithmetic and algebra
  • ↳Trained on reasoning traces
  • ↳Preference-optimised on maths data
02Code preference data

Code

  • ↳Writes code across languages
  • ↳Explains unfamiliar code
  • ↳Reasons about program behaviour
  • ↳Reinforced with dedicated code preference data
03384 px · 256 tok/page

Vision & Documents

  • ↳Describes images and answers questions about them
  • ↳Reasons over visual content
  • ↳Reads scanned pages, PDFs and forms
  • ↳Transcribes charts and screenshots
  • ↳95% content-word recall on OCR
0430 s windows · 24 kHz

Speech & Audio

  • ↳Transcribes speech at 9.7% word error rate
  • ↳Generates natural 24 kHz speech
  • ↳Reasons about spoken input directly
  • ↳Audio feeds the same reasoning core
05512 × 512 · 8 steps

Image Generation

  • ↳Original images from a text prompt
  • ↳8-step LCM sampling, not 30–50
  • ↳Fast enough for interactive use
  • ↳Conditioned by the language core itself
062,097,152 tokens

Scale & Reach

  • ↳2M-token context via Mamba-2 backbone
  • ↳Chunked prefill with state carryover
  • ↳18+ languages, deliberate Hindi strength
  • ↳Deep India knowledge, not a footnote
  • ↳Single-file deployment, one model load
Single Forward Pass

Text, images, documents, speech and audio through one nn.Module, one shared representation space, one model file.