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.
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.
Open knowledge.
Curated open-source models, quantized and made accessible for enterprises and developers. Permissive licenses. HuggingFace-hosted. Drop-in via the EdgeMatrix runtime.
Refined bridges.
Enhanced models — strong open foundations fine-tuned with our datasets to bring sharper performance, trust, and relevance. First Nexon releasing soon.
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
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.
We benchmark every release before it ships. Models that don't hold up don't make the catalog.
Q4_KM, Q5_KM, and Q8 variants where they meaningfully reduce footprint. Same quality bar, smaller binary.
Apache 2.0 / MIT / OpenRAIL where the upstream license allows. No surprise restrictions.
Every Lexicon ships with a manifest the EdgeMatrix runtime understands — load and serve in one line.
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.
Targeted fine-tunes on enterprise corpora — telephony, contracts, claims, code-switched calls. Higher accuracy on the workloads our customers actually run.
Trained alongside HaluMon evaluation. Hallucination rates measured before release. Confidence calibration tuned for regulated deployment.
Indic languages, BFSI compliance vocabulary, healthcare clinical terminology. The vocabulary your customers expect, not what generic instruction-tuning produces.
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.
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.
The current generation of language and vision models has evolved through scale-driven iteration — abundant compute, large memory budgets, cloud-centric inference. Shakti was architected under a different assumption set: compute is constrained, latency matters, and models must adapt to devices, not the other way around.
The objective: maximize intelligence per parameter while preserving predictable behavior under real-world constraints. Every architectural choice — across both Text SLMs and Vision-Language Models — follows from that shift. This is why Shakti models punch 2-3× above their parameter weight.
Query heads dynamically share key/value representations; K/V projections are reused across multiple query heads. Cuts memory-bandwidth pressure, improves cache locality, and — more importantly — measurably improves long-horizon reasoning coherence.
Sliding-window attention, aggressive KV caching, and high-theta RoPE (θ = 500,000). Context is treated as temporal memory — relevance decays gracefully. Reduces hallucination drift in long conversations.
Pre-Norm consistently across layers, with SwiGLU activations. Gradient stability under aggressive fine-tuning; faster convergence on domain adaptation; less sensitivity to prompt phrasing.
48 layers, 1920 hidden dim. Vision tokens treated as reasoning tokens — spatial structure preserved deep into the network, not collapsed by early pooling.
RMSNorm applied inside vision attention. Prevents attention-score explosion on high-resolution images; stabilizes gradients across the deep stack.
14×14 for low-resolution inputs up to 32×32 for high-resolution. Vision adapts to data variability instead of forcing uniformity.
RoPE extended into 2D spatial domains plus absolute positional bias — relative relationships preserved, drift-free grounding in large layouts.
"Trustworthy behavior begins at the architecture layer, not the alignment layer. Alignment techniques refine behavior, but architectural stability determines whether that behavior is reliable under pressure."
Read the full architecture deep-dive →Shakti-2.5B (Q4_KM) benchmarked against Llama 3 8B and Phi-3.5-Mini. Bold = Shakti leads.
Source: Shakti-2.5B technical report — arXiv 2410.11331
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.
Improved stability and convergence behavior.
Pre-LayerNorm early, Post-LayerNorm with RMSNorm later — optimal stability/efficiency balance.
Pre-train, alignment, fine-tune. Lower token budget. Better task generalization.
Source: Shakti-VLM technical report — arXiv 2502.17092