• Konic Uno-1Open-weight family — download, self-host, evaluate.
  • Konic Duo-1Enterprise production family — annual enterprise licence, on-prem.
  • Konic Tres-1Mission-critical tier, highest accuracy bar.
  • All model familiesCompare Uno-1, Duo-1, and Tres-1 — deployment and licence options.
  • Custom LLM DevelopmentLLMs built for your agentic pipelines from company data and requirements.
  • Open Models on HFExplore open-weight releases on Hugging Face.
  • ResearchBenchmarks, compression notes, and release write-ups from the lab.
  • Optimized LFM2.5-VL-3B: FFN Width Pruned, Distillation Aligned, INT4 QuantizedWe surgically compress Liquid AI’s LFM2.5-VL-3B — an architecture-searched hybrid conv+attention vision-language model — by cutting FFN width where the search never optimized it: joint-SwiGLU Wanda pruning (10752→8192 and →7168), distillation-to-baseline LoRA recovery, and hand-rolled W4A16 GPTQ + W8 mixed-precision quantization. All three tiers are near-lossless against the base on our paired eval shards — PPL ratios at or below baseline (0.79–0.85), top-5 token agreement ≥ 0.95, quantization adding only +0.02–0.06 nats — across 5.30 GB (−15%), 4.92 GB (−21%), and 1.93 GB packed / 2.22 GB (Konic Optimized) — native INT4/INT8 compressed-tensors in vLLM (−69%). Live comparison on the smallest tier preserves scenes, facts, and tool calling — 10/10 tool rounds, including correct abstention where the base over-triggered. Recorded runs, not general benchmarks.
  • LFM2.5 Encoder 230M + SigLIP2: A Compact Multimodal EncoderWe augment Liquid AI’s LFM2.5 Encoder 230M — a 230M-parameter bidirectional masked-language encoder — with a SigLIP2 vision tower and 32 learned soft tokens to build an encoder-only multimodal model for retrieval and image-text matching. Clean image-only retrieval on 12,500 held-out MONET pairs reaches 0.1194 image→text R@1 with the BF16 reference; the GPTQ INT4 release retains 0.1091 while shrinking the package from 923.65 MB to 370.46 MB (−59.9%). Cyclic-negative matching AUROC is 0.9733; text-nearest hard negatives and image-conditioned masked-token prediction remain at chance, and the model does not generate.
  • Two Stages, Much Smaller MoE: REAP Expert Pruning Followed by AWQ INT4 QuantizationWe chain two complementary compression stages on Liquid LFM2.5-8B-A1B — REAP CUDA expert pruning (32→16 experts across 22 MoE layers) then external AWQ INT4 quantization (W4A16_ASYM, group 128). The published packed artifact is 1.15B packed-weight equivalent / 2.79 GB; recorded MATH500 and BFCLv3 results use a separate 9.18 GB AWQ-scaled BF16 derivative for vLLM evaluation, not direct packed-INT4 runtime. These are recorded run summaries, not general benchmarks.
  • From-Scratch AWQ INT4 Quantization on Qwen3-8BWe validate a from-scratch, pure-PyTorch AWQ implementation on Qwen3-8B: group-wise INT4 with per-channel AWQ scaling produces a 4.0× smaller model (13.9 GB → 3.5 GB linear weights) at 1.034× FP16 perplexity on WikiText-2 (10.08 vs 9.75), loaded and run in a real INT4 GEMM runtime. The decisive factor is norm-folding the AWQ scale — 20× more accurate per weight than weight-dequantization.
  • On-Prem LLM GuidesDeployment, cost, data sovereignty, and air-gapped LLM guides.
  • LLM Data Sovereignty: Why Enterprises Keep Models In-HouseData sovereignty is the reason most regulated enterprises cannot use hosted LLM APIs: the moment proprietary data crosses a network boundary, control over it is shared. This guide explains what data sovereignty means for AI, why it is driving on-prem adoption, and how to evaluate a deployment against your sovereignty requirements.
  • Air-Gapped LLM Deployment: What It Means and When You Need ItAn air-gapped LLM runs on infrastructure with no connection to the public internet — the model, the data, and the serving stack all live inside your boundary. This guide explains what air-gapped deployment actually requires, which industries need it, and the practical constraints of running a model with no external dependencies.
  • How to Deploy an LLM On-Premise: A Practical GuideDeploying an LLM on your own infrastructure is a sequence of concrete decisions: pick the model, size the hardware, choose a serving runtime, quantize for your GPU, and wire in monitoring. This guide walks each step with the trade-offs that actually matter for a production on-prem deployment.
  • On-Prem LLM vs API: When to Self-Host Your Language ModelsAPI-based LLMs are fast to adopt but scale poorly on cost, latency, and data control. On-prem LLMs trade setup effort for predictable economics, lower per-token cost at volume, and data that never leaves your boundary. This guide breaks down the decision across cost, latency, privacy, compliance, and operational effort — and when each path is the right call.
  • Book a demoTalk with the team about your workload.
  • GitHubOpen research and tooling.
  • Hugging FaceModels and model cards.
  • XFollow @koniclabs on X.
Pricing
Models on HFGitHubX
Book a demo
Book a demo

On-prem LLMs.for enterprise AI.

Book a demoSee models

Konic — enterprise LLM platform for on-prem deployment. Production-optimized enterprise model families (Uno-1 open weights, Duo-1 enterprise production, Tres-1 mission-critical) for enterprise AI workloads in banking, insurance, healthcare, and manufacturing. Annual enterprise licence with no per-token cost: on-prem inference, private cloud VPC, air-gapped and sovereign deployment on infrastructure you own. Custom enterprise LLM development from company data. NVIDIA Inception Program member. Research: konic.io/research.

We help enterprises build

production LLMs — Konic model family,

and custom models for their workflows.

Konic ModelsCustom development
Research

Optimized LFM2.5-VL-3B: FFN Width Pruned, Distillation Aligned, INT4 Quantized

We surgically compress Liquid AI’s LFM2.5-VL-3B — an architecture-searched hybrid conv+attention vision-language model — by cutting FFN width where the search never optimized it: joint-SwiGLU Wanda pruning (10752→8192 and →7168), distillation-to-baseline LoRA recovery, and hand-rolled W4A16 GPTQ + W8 mixed-precision quantization. All three tiers are near-lossless against the base on our paired eval shards — PPL ratios at or below baseline (0.79–0.85), top-5 token agreement ≥ 0.95, quantization adding only +0.02–0.06 nats — across 5.30 GB (−15%), 4.92 GB (−21%), and 1.93 GB packed / 2.22 GB (Konic Optimized) — native INT4/INT8 compressed-tensors in vLLM (−69%). Live comparison on the smallest tier preserves scenes, facts, and tool calling — 10/10 tool rounds, including correct abstention where the base over-triggered. Recorded runs, not general benchmarks.

NVIDIA Inception Program member badgeAnnouncement (LinkedIn)
Latest news

Konic joins the NVIDIA Inception Program

Company

Open-sourcing Agentbox — Docker-based RL training agent runtime

Open source

Konic joins the NVIDIA Inception Program

Do it all with Konic.

Open models.

Explore Uno-1

Published openly for developer adoption and ecosystem reach — compact weights you can run, fork, and ship.

Explore Uno-1
Open model registry with model families and deployment metadata
  • Open
  • Flexible
  • Developer-ready
  • Deployable
  • Community-driven

Production LLMs.

See Duo-1

Balanced capability for production workloads — the cost-effective path for specialized language models you control.

See Duo-1
Production language model observability console
  • Production-ready
  • Reliable
  • Efficient
  • Scalable
  • Integration-friendly

Mission-critical accuracy.

See Tres-1

Highest capability where accuracy carries the most risk — specialized models for regulated and high-stakes work.

See Tres-1
Mission-critical model evaluation and audit workspace
  • Accurate
  • Robust
  • Secure
  • Auditable
  • Enterprise-ready

Open research.

Read research

Compression, quantization, and pruning results published openly — methods that make frontier capability fit production hardware.

Read research
Open research methods connected through a model research map
  • Open
  • Experimental
  • Reproducible
  • Collaborative
  • Forward-looking

Infrastructure you control.

Book a demo

Deployed inside your environment, on infrastructure you own — versioned releases without changing your integration.

Book a demo
Private infrastructure capacity and deployment topology
  • Private
  • Controlled
  • Portable
  • Secure
  • Operationally ready

Built with partners.

NVIDIA Inception

NVIDIA Inception member — production AI on modern GPU stacks, with research that ships into real systems.

NVIDIA Inception
Konic deployment connected to infrastructure and research partners
  • Connected
  • Supported
  • Scalable
  • Enterprise-ready
  • Collaborative
Konic

Compact production-optimized LLMs on infrastructure you own.

NVIDIA Inception Program memberAnnouncement (LinkedIn)

Models

  • Konic Models
  • Custom LLM Development
  • Pricing
  • Open Models

Research

  • How We Build
  • About Konic
  • Glossary
  • Research
  • Optimized LFM2.5-VL-3B: FFN Width Pruned, Distillation Aligned, INT4 Quantized
  • LFM2.5 Encoder 230M + SigLIP2: A Compact Multimodal Encoder
  • Two Stages, Much Smaller MoE: REAP Expert Pruning Followed by AWQ INT4 Quantization

Guides

  • On-Prem LLM Guides
  • LLM Data Sovereignty: Why Enterprises Keep Models In-House
  • Air-Gapped LLM Deployment: What It Means and When You Need It
  • How to Deploy an LLM On-Premise: A Practical Guide

Connect

  • Book a demo