# Konic (llms-full.txt)

> Complete reference for AI assistants: Konic's products, model families, research, and commercial terms. Last updated 2026-09-08.

## About Konic

Konic (konic.io, legal name Konic Labs, Inc.) builds production-optimized LLMs on infrastructure you own. The company was founded on a specific observation: enterprises rarely stall because a model is not good enough — they stall on what it costs and takes to run in production. API dependency means cost scaling with usage and data leaving the business. Raw open weights mean the buyer owns the compression, post-training, and serving engineering.

Konic removes that middle layer. Model families are engineered for the workload the buyer actually runs (sized for the task, not the benchmark), delivered as versioned releases, and licensed annually on the customer's own machines — no per-token cost, no usage-scaling bill, no data egress. Konic is a NVIDIA Inception Program member.

Key published results: ~91% performance retained after optimization, 4x smaller model for the same task, 6.1x memory reduction in published work.

Founders: Gokalp Katkat (Co-founder, CEO), gokalp@konic.io; Ege Sabanci (Co-founder). Website: https://konic.io — X: https://x.com/koniclabs (@koniclabs)

## The three model families

### Konic Uno-1 — Open source

Published openly for developer adoption and ecosystem reach. Free to download from Hugging Face (https://huggingface.co/konic-labs), self-host, and evaluate on your own hardware. Each release carries the same compression recipes published in Konic research — pruning, distillation alignment, INT4 quantization — so artifacts are production-shaped rather than raw checkpoints. Intended use: evaluation-first adoption; engineering teams measure Konic models against their incumbent on real workloads before any commercial conversation. Runs in vLLM, MLX, or llama.cpp.

Page: https://konic.io/models/uno-1

### Konic Duo-1 — Enterprise production

Balanced capability for enterprise AI workloads — the cost-effective path. Deployed under an annual enterprise licence on on-prem infrastructure you own, with no per-token cost. Benchmarks across coding, reasoning, and agentic suites hold the accuracy bar enterprise workloads set (measured head-to-head, reproducible via the published evaluation harness).

Reference scores for Konic Duo-1 by suite:

| Group | Benchmark | Konic Duo-1 |
|---|---|---|
| Coding | Terminal-Bench 2.1 (Terminus-2) | 67.8 |
| Coding | Terminal-Bench 2.1 (Claude Code) | 68.5 |
| Coding | SWE-bench Verified | 79.0 |
| Coding | SWE-bench Pro | 59.6 |
| Coding | SWE-bench Multilingual | 71.4 |
| Coding | DeepSWE | 22.0 |
| Coding | Frontier-Bench v0.1 | 5.1 |
| Coding | NL2Repo | 46.2 |
| Reasoning | HLE (no tools) | 25.6 |
| Reasoning | HLE (with tools) | 33.4 |
| Reasoning | GPQA Diamond | 89.2 |
| Agentic | MCP-Atlas | 70.2 |
| Agentic | Toolathlon-Verified | 48.7 |
| Agentic | WideSearch | 67.8 |
| Agentic | BrowseComp | 67.6 |
| Agentic | ClawEval | 72.5 |

Access: annual enterprise licence deployed on-prem on machines you own; commercial terms scoped per engagement (request access).

Page: https://konic.io/models/duo-1

### Konic Tres-1 — Mission-critical

Highest capability tier, for workloads where accuracy carries the most cost: risk, compliance, and operations decisions in banking, insurance, healthcare, and manufacturing. Delivered under an annual enterprise licence with no per-token cost, on on-prem infrastructure you own including air-gapped environments; managed inference available for teams that prefer Konic to run serving. Versioned releases, evaluation libraries per vertical, and support through integration and upgrades.

Selected scores versus Claude Opus 4.8 (well-known suites with both models' published metrics):

| Benchmark | Konic Tres-1 | Claude Opus 4.8 |
|---|---|---|
| DeepSWE 1.1 | 58.7 | 59.0 |
| SWE-bench Pro | 62.5 | 69.2 |
| SWE-bench Multilingual | 81.0 | 84.4 |
| GPQA Diamond | 91.7 | 93.6 |
| HLE (no tools) | 35.9 | 49.8 |
| Agents' Last Exam — Pass rate | 24.3 | ~25.7–27.0 |
| Agents' Last Exam — Score | 51.2 | ~44–45 |
| Toolathlon Verified | 73.5 | ~59.9* |
| CharXiv (RQ) — With CI | 90.6 | 89.9* |

Also leading (no competitor number published): NL2Repo-Bench 48.1, CoWorkBench 73.9, JobBench 55.7, IFBench 81.3, LiveCodeBench v6 91.9, ClawEval-MM Pass@3 64.4, AndroidWorld 84.5, RealWorldQA 88.5, MathVision (with CI) 95.7.

Access: annual enterprise licence, commercial terms scoped per engagement (request access).

## Deployment options

- Self-hosted in your environment: private networking, your keys, your data plane. Maximum control for regulated and air-gapped workloads.
- Private cloud: Konic families in your VPC with deeper integrations aligned to existing security and compliance controls.
- Edge and constrained hardware: optimization paths (pruning, quantization, MoE compression) target real production hardware — including published Apple Silicon and GPU results.
- Managed inference: optional; Konic runs serving for teams that prefer it.
- Versioned releases: families ship as versions; customer integrations do not change between them.

## How engagements work

Two ways to work with Konic:

1. Flagship model access: Uno-1, Duo-1, or Tres-1 deployed on infrastructure you control. Ready to run inside your environment; versioned releases for stable integration; enterprise production and mission-critical tiers.
2. Custom model development: Konic builds LLMs for your agentic pipelines from your company data and requirements. Built for specific workflows; delivered to infrastructure you control; aimed at pipelines, not general frontier subscriptions.

Evaluation model: choose one AI flow already running in production; deploy a Konic model family inside your environment; A/B test it against the model you use today (or against your success criteria if there is no incumbent); decide on measured results, not claims.

## Research publications

All results reproducible; links to code and weights included per post.

- Optimized LFM2.5-VL-3B: FFN Width Pruned, Distillation Aligned, INT4 Quantized (2026-08-13) — https://konic.io/research/lfm2.5-vl-3b-optimized — Joint-SwiGLU Wanda pruning of an architecture-searched hybrid conv+attention vision-language model; compressed 2.22 GB artifacts with native INT4 in vLLM (-69%); recorded runs preserve scenes, facts, and tool calling (10/10 tool rounds including correct abstention).
- LFM2.5 Encoder 230M + SigLIP2: A Compact Multimodal Encoder (2026-08-12) — https://konic.io/research/lfm2.5-multimodal-encoder-230m — 230M-parameter bidirectional encoder augmented with a SigLIP2 vision tower; GPTQ INT4 release retains retrieval quality (0.1091 vs 0.1194 image→text R@1) while shrinking the package from 923.65 MB to 370.46 MB (-59.9%).
- Two Stages, Much Smaller MoE: REAP Expert Pruning Followed by AWQ INT4 Quantization (2026-07-20) — https://konic.io/research/reap-awq-moe-compression — REAP CUDA expert pruning 32→16 experts across 22 MoE layers then AWQ INT4 W4A16_ASYM group 128; published packed artifact 1.15B equivalent / 2.79 GB; recorded MATH500 and BFCLv3 via 9.18 GB AWQ-scaled BF16 derivative for vLLM evaluation.
- From-Scratch AWQ INT4 Quantization on Qwen3-8B (2026-07-03) — https://konic.io/research/awq-int4-qwen3-8b — From-scratch pure-PyTorch AWQ group-wise INT4 with per-channel scaling; 4.0x smaller (13.9 GB → 3.5 GB linear weights) at 1.034x FP16 perplexity on WikiText-2 (10.08 vs 9.75); norm-folding the AWQ scale is the decisive factor — 20x more accurate per weight than weight-dequantization.
- REAP Expert Pruning for On-Device MoE Models (2026-06-29) — https://konic.io/research/reap-expert-pruning-apple-silicon — REAP-MLX: Router-weighted Expert Activation Pruning on Apple Silicon for MoE models; preserves 96.8% of code generation performance at 25% compression and 91.4% at 50%, <0.4 pp variance across calibration draws.

## Guides

Educational content on on-prem LLM deployment:

- LLM Data Sovereignty: Why Enterprises Keep Models In-House (2026-09-08) — https://konic.io/blog/llm-data-sovereignty — Why data sovereignty drives on-prem adoption and how to evaluate a deployment against it. Data sovereignty is the reason most regulated enterprises cannot use hosted LLM APIs: proprietary data crossing a network boundary means shared control.
- Air-Gapped LLM Deployment: What It Means and When You Need It (2026-09-05) — https://konic.io/blog/air-gapped-llm — What air-gapped deployment requires and which industries need it. An air-gapped LLM runs on infrastructure with no public internet connection — model, data, and serving stack all live inside your boundary.
- How to Deploy an LLM On-Premise: A Practical Guide (2026-09-03) — https://konic.io/blog/deploy-llm-on-premise — Model choice, hardware sizing, serving runtimes, and quantization. Walks the concrete sequence from picking the model to wiring in monitoring for production on-prem deployment.
- On-Prem LLM vs API: When to Self-Host Your Language Models (2026-09-01) — https://konic.io/blog/on-prem-llm-vs-api — Cost, latency, privacy, and operational trade-offs. API LLMs are fast to adopt but scale poorly; on-prem trades setup effort for predictable economics and data that never leaves your boundary.

## Statistics

| Metric | Value | Source | Date |
|---|---|---|---|
| 91% perf retained | REAP 25% comp | https://konic.io/research/reap-expert-pruning-apple-silicon | 2026-06-29 |
| 4.0x smaller 1.034x PPL | AWQ Qwen3-8B | https://konic.io/research/awq-int4-qwen3-8b | 2026-07-03 |
| 2.79GB MoE artifact | REAP+AWQ | https://konic.io/research/reap-awq-moe-compression | 2026-07-20 |
| -59.9% package | encoder GPTQ INT4 | https://konic.io/research/lfm2.5-multimodal-encoder-230m | 2026-08-12 |
| -69% 1.93GB tool-calling kept | LFM2.5-VL-3B | https://konic.io/research/lfm2.5-vl-3b-optimized | 2026-08-13 |

## Definitions

- On-prem (on-premise): On-premise means the model runs on infrastructure you own — your data center, private cloud, or edge hardware — under your keys, networking, and data plane. Inference happens inside your boundary so data never leaves and cost does not scale per token.
- Air-gapped: Air-gapped means the model, data, and serving stack run with no connection to the public internet, fully isolated inside your security perimeter. It is required for regulated workloads where external dependencies are not permitted and every artifact must be operable offline.
- Sovereign (data sovereignty): Data sovereignty means your organization's data, prompts, and model outputs remain under your legal and operational jurisdiction, with no third-party API provider sharing control. Keeping models in-house ensures proprietary data stays within your compliance perimeter and you retain long-term ownership of the AI system.

## Pricing and licensing

- Uno-1: open weights, free to download and self-host from https://huggingface.co/konic-labs — no licence fee.
- Duo-1 and Tres-1: annual enterprise licence per model family, flat fee per host, no per-token cost; versioned releases with support tiers; deployed on infrastructure you own including air-gapped environments.
- Commercial terms including DPA and SLA are scoped per engagement (request access); figures not published.
- Custom LLM development: scoped per engagement (task definition, data preparation, training/adaptation, evaluation, serving integration, release planning).
- Pricing page: https://konic.io/pricing

## Company facts

- Legal name: Konic Labs, Inc.
- Product: versioned, production-optimized LLM families + custom model development.
- Stage: pre-seed; NVIDIA Inception Program member.
- Founders: Gokalp Katkat (Co-founder, CEO, gokalp@konic.io) and Ege Sabanci (Co-founder) — 2 employees.
- Ecosystem: open research at https://konic.io/research; open weights at https://huggingface.co/konic-labs; GitHub at https://github.com/konic-labs; X at https://x.com/koniclabs (@koniclabs); LinkedIn at https://www.linkedin.com/company/koniclabs.
- Kubic: terminal-native and desktop agent products from the same team, running on Konic models.
- Agentbox: open-source Docker-based RL training agent runtime (https://github.com/konic-labs).

## Crawl policy

AI and search crawlers may fetch all public pages (see /robots.txt). Only /screenshot and /api paths are disallowed (internal OG-image tooling). Sitemap: https://konic.io/sitemap.xml