Open-Weight LLM · Private & Custom AI
Ornith-1.0-35B
Agentic coding model optimized for self-hosted deployment; built to automate software engineering tasks—code generation, debugging, repository navigation—within your own infrastructure.
Ornith-1.0-35B is a 35B-parameter mixture-of-experts model post-trained on Qwen 3.5, specialized in coding-agent reasoning (SWE-Bench, Terminal-Bench, NL2Repo). It's designed for single-GPU deployment without regional restrictions. For ops teams, this means a private, fully-controlled LLM backbone for building internal developer tooling, code automation, and knowledge-retrieval agents that never leave your network.
Model facts
Private deployment
Run Ornith-1.0-35B in your own environment
Ornith-1.0-35B can run on a single high-memory GPU (~80GB VRAM at FP16, ~40GB quantized). Being MIT-licensed and non-gated, you can download, host, and operate it entirely within your data center or private cloud with no external API calls or vendor lock-in. All code-generation outputs and intermediate reasoning stay in your environment—critical for teams handling proprietary codebases or IP-sensitive development.
Operational AI use cases
Internal Code Review & Compliance Automation
Deploy Ornith as a self-hosted agent that scans pull requests, flags security patterns, enforces internal coding standards, and drafts review comments. Runs entirely on-premises; integrates with GitLab/GitHub webhooks. No code or diffs ever reach external APIs.
Repository Navigation & Developer Onboarding
Build a private RAG layer on top of Ornith to index your codebase. New engineers ask questions about repo structure, dependencies, and patterns; the model retrieves context from local git/docs and generates answers. Reduces onboarding friction while keeping code private.
Bug Triage & Root-Cause Analysis
Ingest error logs, stack traces, and issue descriptions into a local agent powered by Ornith. Model generates hypotheses, suggests test cases, and links to relevant code sections. Operationalizes your support/QA workflow without exposing incident data.
Custom AI
As a base for custom AI
Strong fit for building proprietary developer-productivity applications. Use Ornith as the base LLM in a custom IDE plugin, internal CLI tool, or ticket-to-code workflow. Fine-tune on your specific codebase patterns, architecture, and company conventions. MIT license permits commercial products built on top; you retain full IP ownership.
In the operating system
Where it fits
Operates as the **agent reasoning layer** in an AI operating system for engineering teams. Sits above a code-indexing/knowledge layer (local vector DB of repos, docs, runbooks) and below orchestration (task scheduling, tool-use endpoints). Can power autonomous agents that invoke build systems, test runners, and deployment pipelines.
Data control & security
Self-hosting eliminates data exfiltration risk. Code, diffs, commit history, and internal documentation never leave your network. No telemetry or model-improvement feedback loop tied to external parties. Note: security posture depends on your infrastructure hardening; the model itself is not a security mechanism. Compliance benefits (e.g., data residency for regulated industries) derive from the *deployment architecture*, not the model weights.
Hardware footprint
**Estimate (unverified)**: ~80GB VRAM (FP32), ~40GB (FP16), ~20GB (int8 quantized), ~10–12GB (int4 quantized). Single high-end GPU (e.g., A100 80GB, L40, RTX 6000) sufficient for inference; multi-GPU setup for high throughput or fine-tuning. Quantization recommended for cost/latency trade-off.
Integration
Supports transformers/HuggingFace inference libraries (vLLM, TGI, Ollama). Accepts text and image-text inputs. Stateless API—integrate via REST/gRPC into CI/CD pipelines, chat interfaces, or async job queues. Batch inference possible; latency depends on quantization and hardware. Context length unknown; verify max-token handling before wiring into multi-turn agent workflows.
When it's not the right fit
- —Context length is not specified in the model card. If you need 10K+ token windows for long-document reasoning or multi-file code understanding, verify compatibility first.
- —Non-coding reasoning tasks (general knowledge QA, creative writing, reasoning over unstructured prose). Ornith is heavily tuned for code; general-purpose tasks may underperform.
- —Fine-tuning resources are scarce. Model card does not detail how to adapt Ornith for your domain; you'll need to conduct your own training experiments and benchmarking.
- —Compliance/audit trails for model behavior are not documented. If your workflow requires interpretability logs or deterministic outputs for regulatory sign-off, you'll build that yourself.
Alternatives to consider
DeepSeek-Coder-33B
Similar size, strong code performance, permissive license. Slightly different training approach; compare SWE-Bench scores for your use case.
Qwen 3.5-35B (base model)
Ornith is fine-tuned on this; if you need broader reasoning or lower cost, use Qwen directly and fine-tune yourself. Less coding optimization.
Llama 3.1-70B
Larger, broader capability. Stronger general reasoning, more docs, bigger community. Requires more GPU VRAM; coding performance may lag Ornith.
Related open models
FAQ
Can we run Ornith-35B in our private cloud without touching external APIs?
Yes. The model is MIT-licensed, non-gated, and compatible with standard inference frameworks (vLLM, TGI, Ollama). Download weights once, deploy on your infrastructure. No vendor call-home or telemetry required.
Can we use this in a commercial product we sell to customers?
Yes. MIT license permits commercial use, modification, and distribution. You own the resulting product. You must include the license notice. No royalties or restrictions.
What's the max context window?
Unknown. Model card does not specify. Check HuggingFace model config or test empirically before deploying to long-context workflows.
How does Ornith handle image inputs?
Tags indicate image-text-to-text capability, but the model card does not detail format, resolution limits, or performance. Requires hands-on testing or engineering deep-dive.
Build proprietary AI agents on Ornith—fully private, fully yours.
LLM.co helps you deploy open-weight LLMs like Ornith-35B into your infrastructure, integrate them into ops workflows, and fine-tune for your codebase. Skip the API tax. Control your data. Let's architect your AI OS.