// L04 — FOUNDATION MODELS

A continuum of intelligence.
Open knowledge, refined bridges,
original innovation.

Three model lines that work together. Lexicons — curated open-source models, quantized for enterprise. Nexons — open foundations enhanced with our own datasets for sharper performance, trust, and relevance. Shakti — a fully in-house family of small and mid language models, built ground-up for edge and enterprise AI.

Shakti released
6
Largest in flight
30B
arXiv papers
3
Patent filings
PCT

From open knowledge to original innovation.

Most enterprises don't need a single model — they need the right model for the workload, with a path to upgrade as their data matures. Lexicons, Nexons, and Shakti are the three rungs of that ladder: from accessible open-source baselines, through enhanced fine-tunes, to fully in-house frontier work.

[ PILLAR 01 ]

Lexicons

Open knowledge.

Curated open-source models, quantized and made accessible for enterprises and developers. Permissive licenses. HuggingFace-hosted. Drop-in via the EdgeMatrix runtime.

[ PILLAR 02 ]

Nexons

Refined bridges.

Enhanced models — strong open foundations fine-tuned with our datasets to bring sharper performance, trust, and relevance. First Nexon releasing soon.

[ PILLAR 03 ]

Shakti

Original innovation.

A fully in-house family of small and mid language models, built ground-up for edge and enterprise AI. Six released (100M – 4B). Two in flight (8B, 30B). Three arXiv papers.

"From open knowledge, to refined bridges, to original innovation. The vision: make AI accessible, reliable, and meaningful — whether it runs on the cloud, on-premise, or at the edge."

— Kamalakar Devaki, Founder · CEO, SandLogic Technologies

Open-source models, made enterprise-ready.

Lexicons is our growing zoo of curated open-source foundation models — quantized, packaged, and benchmarked for enterprise deployment. Permissive licenses on HuggingFace and GitHub. Quick customization, minimal retraining, full transparency.

Curated

We benchmark every release before it ships. Models that don't hold up don't make the catalog.

Quantized

Q4_KM, Q5_KM, and Q8 variants where they meaningfully reduce footprint. Same quality bar, smaller binary.

Permissive

Apache 2.0 / MIT / OpenRAIL where the upstream license allows. No surprise restrictions.

Runtime-ready

Every Lexicon ships with a manifest the EdgeMatrix runtime understands — load and serve in one line.

$ pip install sandlogic-lexicons
# Browse the catalog
from lexicons import catalog
print(catalog.list(domain="finance"))
# Load any model
from lexicons import load
model = load("shakti-2.5b-q4km")
# Or run via EdgeMatrix runtime
from edgematrix import Runtime
rt = Runtime("shakti-vlm-4b", device="cuda")
output = rt.generate(prompt, image=img)

Open foundations, sharpened with our data.

Nexons take strong open foundations and fine-tune them with SandLogic's proprietary datasets to bring sharper performance, trust, and relevance for enterprise workloads. The bridge between the open ecosystem and the in-house Shakti family — built for teams that want better-than-baseline accuracy without committing to a fully proprietary stack.

Sharper performance

Targeted fine-tunes on enterprise corpora — telephony, contracts, claims, code-switched calls. Higher accuracy on the workloads our customers actually run.

Trust by construction

Trained alongside HaluMon evaluation. Hallucination rates measured before release. Confidence calibration tuned for regulated deployment.

Domain relevance

Indic languages, BFSI compliance vocabulary, healthcare clinical terminology. The vocabulary your customers expect, not what generic instruction-tuning produces.

[ COMING SOON ]

First Nexon releases shortly.

The first Nexon is in late-stage evaluation now. Sign up to be notified when it ships — or talk to us if you want a Nexon trained on your domain corpus before the public release.

One family. Every parameter range.

Shakti is our fully in-house family of small and mid language models — built ground-up for edge and enterprise AI. Six released (100M to 4B parameters), two in flight (8B and 30B). Pick the smallest model that meets your accuracy bar — Shakti models are tuned to outperform peers 2–3× their size, so the right deployment is almost always smaller than you think.

Shakti-100M
100M
Always-on wearables, IoT, ultra-low-power
Optimized for IoT command grammars
Shakti-250M
250M
Smartwatch, hearables, mid-tier embedded
Outperforms SmolLM-135M on Anli & Piqa
Shakti-500M
500M
Smart home, voice assistants, edge inference
Optimized RLHF + DPO alignment
Shakti-1B
1B
Document intelligence, OCR, multimodal
OCRBench: 798 · Beats InternVL2-1B, MiniCPM-V-2
Shakti-2.5B
2.5B
Enterprise agents, contact center, BFSI
MMLU 69.2 (Q4_KM) · Beats Phi-3.5-Mini · arXiv published
Shakti-4B-VLM
4B
High-end document VLM, multi-modal reasoning
DocVQA 92.92 · Beats InternVL2-4B, Phi-3-Vision
Shakti-8B
8B
Scientific research, comprehensive analysis
In flight — extended context, domain-specific
Shakti-30B
30B
Frontier reasoning at the edge of efficiency
In flight — H2 2026
// PARAMETER SCALE

From wearables to frontier — log-scaled.

100MShakti-100M250MShakti-250M500MShakti-500M1BShakti-1B2.5BShakti-2.5B4BShakti-4B-VLM8BShakti-8B30BShakti-30BShippedIn flight
Bar height represents log-scaled parameter count. Hatched bars are models in flight (8B, 30B).
// SHAKTI-2.5B vs PEERS

3× smaller.
Match for match.

Shakti-2.5B (Q4_KM) benchmarked against Llama 3 8B and Phi-3.5-Mini. Bold = Shakti leads.

MMLU (5-shot)
Llama 3 8B: 58.1
69.2
PIQA (5-shot)
Phi-3.5-Mini: 78.0
84.7
WinoGrande (5-shot)
Phi-3.5-Mini: 65.1
68.1
SocialQA
Llama 3 8B: 71.4
78.1
MedQA
Phi-3.5-Mini: 47.4
57.1
TruthfulQA (MC2)
Llama 3 8B: 55.1
61.2
// SHAKTI-2.5B vs LLAMA 3 8B (selected benchmarks)

MMLU · SocialQA · TruthfulQA — head to head.

MMLU (5-shot)Shakti-2.5B Q4_KM69.2MMLU (5-shot)Llama 3 8B58.1SocialQAShakti-2.5B78.1SocialQALlama 3 8B71.4TruthfulQA (MC2)Shakti-2.5B61.2TruthfulQA (MC2)Llama 3 8B55.1
Higher is better. Methodology in arXiv 2410.11331.

Source: Shakti-2.5B technical report — arXiv 2410.11331

Vision-language at a fraction of the size.

Shakti-VLM-4B uses QK-normalization and hybrid normalization (Pre-LayerNorm in early layers, Post-LayerNorm with RMSNorm in later ones). Despite using significantly fewer training tokens, it beats Qwen2VL-7B and MiniCPM-V-2.6-8B on document and chart understanding.

MMMU (val)
59.78
Phi-3-Vision: 46.1
DocVQA (test)
92.92
InternVL2-4B: 89.2
OCRBench
849
Phi-3-Vision: 639
TextVQA (val)
85.56
InternVL2-4B: 74.4
// SHAKTI-VLM-4B vs PEERS

Document understanding — at 4B parameters.

DocVQA (test)Shakti-VLM-4B92.92DocVQA (test)InternVL2-4B89.2TextVQA (val)Shakti-VLM-4B85.56TextVQA (val)InternVL2-4B74.4MMMU (val)Shakti-VLM-4B59.78MMMU (val)Phi-3-Vision46.1
Higher is better. Sources: Shakti-VLM technical report — arXiv 2502.17092.

QK-Normalization

Improved stability and convergence behavior.

Hybrid normalization

Pre-LayerNorm early, Post-LayerNorm with RMSNorm later — optimal stability/efficiency balance.

Three-stage training

Pre-train, alignment, fine-tune. Lower token budget. Better task generalization.

Source: Shakti-VLM technical report — arXiv 2502.17092

// LET'S BUILD

Pick the model that fits your problem.