Skip to main content

Hands-on demos of Squirro Neo, GraphRAG and our agents. No sign-up, no sales call. Try Them Now

Book a Demo

IT · Application Development · All Industries

Your Code Can’t Leave the Organization. Your Developers Still Deserve Frontier-Class AI.

Regulation, data residency, and sovereignty rules keep source code away from cloud-hosted AI, which leaves developers in regulated organizations without the tools the rest of the market already uses. The RMJ2 Coding Assistant powered by Squirro enables open-weight models inside infrastructure you control, in the tools your developers already know, grounded in your own applications and data.

Book a Discovery Call
  • Reduce Risk
  • Phase 1 Available Now
  • Low Effort

Your infrastructure

Trusted by Industry Leaders

library of congress OCBC Bank Zwickroell Security Benefit Henkel Siemens Bertelsmann pavlov media long beach container terminal Deutsche Bundesbank Candriam IDB Invest Bühler logo swissrockets Customer Logo

Measurable Impact

Within 3 Points of the Frontier at Its Best, for Up to 30 Times Less per Run.

  • 86 vs. 89

    out of 100 against Claude Opus 5.5 at maximum effort, on a four-step developer task in a mainframe application

  • About10×

    lower cost on a full application build, delivered in two-thirds of the time

  • +15

    points for the same open model once it works with Squirro (71 to 86), with run-to-run spread down from ±8.3 to ±1.1

Inside Your Boundary. Source code, models, and data stay in your approved infrastructure.

Frontier model against open-weight models with Squirro, same tasks Claude Opus 5.5 Frontier GLM-5.3 with Squirro Open-weight LLM
Model Quality Cost Run time
Four-step developer task in a mainframe application, Opus at maximum effortQuality score out of 100
Claude Opus 5.5 89 $9.10 per run 26–39 min
GLM-5.3 with Squirro 86 $1.70 per run 15–18 min
Same task, lower-cost open modelQuality score out of 100
Claude Opus 5.5 89 $9.10 per run 26–39 min
DeepSeek V4.1 Flash with Squirro 86 $0.29 per run 23–26 min
Full application buildWeighted score out of 10
Claude Opus 5.5 8.6 $65–102 3 h 43 min
GLM-5.3 with Squirro 7.5 $6.95 2 h 29 min

RMJ2 test runs, September–October 2026. Developer task: AWS CardDemo mainframe application, more than 100 runs scored by an AI reviewer against a reference answer, preliminary until manual review is complete. Application build: seven-factor reviewer scorecard. Costs at list prices. Run times vary with provider load.

What It Does

Open-Weight Coding Assistant, Hosted Inside Your Boundary.

Public coding assistants run your code on someone else’s infrastructure, on closed models whose weights, training data, and update cadence you cannot control. For banks, telecoms, and public-sector bodies, that rules them out.

Squirro + RMJ2 pairs selected open-weight models and open-source developer tooling with the RMJ2 Coding Assistant, on top of the Squirro knowledge platform. Developers get the experience they expect. Code, models, and data stay inside the approved boundary.

  • Deploys Where You Need It

    In-country, in a private cloud, on-premises, or air-gapped.

  • Full Modern Toolset

    Code generation and completion, chat, debugging, refactoring, test generation, multi-file changes, agentic tasks, and Git workflows, in the terminal, desktop app, or IDE.

  • Benchmarked Against the Frontier

    Within 3 points of Claude Opus 5.5 at maximum effort on a four-step developer task, and level with Claude Sonnet 5.5 at its maximum setting (86), with code quality on par with flagship models.

  • Model-Agnostic and Governed

    Models validated for quality, latency, security, and licensing, and upgraded as stronger open models appear.

  • Data Sovereignty

    No code or data goes to a model provider.

  • Strategic Autonomy

    Continuity that does not depend on a foreign AI provider’s decisions or service withdrawals.

Customer Proof

I didn’t expect this quality from using GLM with the Squirro coding assistant. I tried to find someone in the market who has an AI coding assistant that’s running on-prem, but I didn’t find anyone with this quality. And I do understand the quality can be improved further with knowledge graphs. This is just the beginning.

Customer feedback after a live demoEnterprise IT solutions provider

The demo ran on GLM-5.3, an open-weight model, in a customized OpenCode terminal and desktop app with VS Code autocomplete. The customer’s developers now use that test environment on their own. The same setup moves to customer-hosted models, optionally air-gapped, with no change to how developers work.

  • Open-weight model
  • Familiar tools
  • No frontier model involved

Customer quote from a live demo, October 2026. Customer name withheld.

How It Works

A Coding Assistant Is Only as Good as Its Model, Its Tools, and What It Knows About Your Code.

Your code never leaves your environment. Squirro + RMJ2 runs the model there, delivers it through tools developers already use, and connects it to the Squirro knowledge graph, so each request carries knowledge about your applications instead of piles of raw files.

  • Models

    Open-Weight Models on Your Infrastructure

    GLM-5.3 is the primary coding model today, with DeepSeek and Qwen models for autocomplete and fallback, hosted in-country, in a private cloud, on-premises, or air-gapped.

  • Tools

    The Tools Developers Already Use

    A customized OpenCode terminal and desktop app, plus VS Code with fast autocomplete, so there is no new interface to learn.

  • Knowledge

    A Knowledge Graph Behind Every Request

    It connects code, documentation, tickets, standards, and dependencies into one queryable model of your estate, so every answer is grounded in your own services and patterns rather than generic training data – with fewer steps, fewer tokens, and less GPU memory per request.

  • Benchmark

    Benchmarked Before Go-Live

    Every model and configuration is tested against frontier models on representative tasks, and in the full application build, reviewers ranked GLM-5.3 with Squirro second to Claude Opus 5.5 overall and first on cost.

Your infrastructure

In-country · Private cloud · On-premises · Air-gapped

Developer tools

OpenCode terminalDesktop appVS Code

Squirro knowledge graph

CodeDocsTicketsStandardsDependencies

Open-weight models

GLM-5.3DeepSeekQwen

Model providersNo code or data goes out

Indicative sizing

  • 1 node8× NVIDIA H200
  • 300–400developers served
  • 35+tokens per second per session, the target
How four models scored on the full application build
Factor (score of 10) Claude Opus 5.5GLM-5.3 with SquirroGrok 4.7Qwen3.6
Adherence to instructions 10 8 6 4
Design and documentation quality 9 8 6 6
Code quality 9 7 6 4
Automated testing (strategy, validity, bugs caught) 10 8 4 5
Maintainability and complexity 9 7 7 4
Cost of implementation 4 9 7 5
Average cyclomatic complexity (lizard) 2.7 3.5 3.2 3.1
Functions above complexity 15 8 of 584 4 of 181 1 of 137 3 of 232

Full application build, October 2026. All four averages sit in the 1–10 “simple, low risk” band. GLM-5.3’s average is raised by a single function, deliberately left uncorrected.

Your Expansion Path.

Each stage adds knowledge about your applications and extends what the assistant can do.

  1. To start

    Sovereign Coding Assistant

    Open-weight model, familiar tools, hosted inside your boundary. Code generation, debugging, refactoring, testing, and agentic tasks from day one.

    Try it in our test environment now

  2. Next

    Enterprise-Aware Coding

    The knowledge graph connects repositories, documentation, APIs, standards, and tickets. A developer asks for a feature that uses your existing authentication service and internal standards, and gets an answer grounded in how your organization builds. Most of the remaining gap to frontier models is missing context, which the knowledge graph supplies.

  3. Next

    Engineering Intelligence

    Cross-application understanding, specification recovery for undocumented systems, and change-impact analysis before a change ships, extended to architects, testers, and analysts.

  4. Vision

    Full Application Intelligence Platform

    A continuously learning model of your application estate, spanning AI coding, knowledge preservation, modernization, and autonomous issue resolution.

Start Where the Friction Is.

The friction looks different in every organization: frontier models that are off-limits for your code, AI coding bills that grow with every developer, and existing applications that generic assistants cannot understand. We are mapping the path to value with a handful of technology teams.

Share your details for early access to the test environment, a look at how this fits your stack, or a say in how it gets built. A few questions first. Then, if it makes sense for both sides, a 20-minute conversation. No pitch.