DEV.co
Open-Source LLM · deepreinforce-ai

Ornith-1.0-397B

Ornith-1.0-397B is a 397-billion-parameter open-source language model optimized for agentic coding tasks. Built on Qwen 3.5 and Gemma 4 foundations, it uses a mixture-of-experts architecture and reinforcement learning to improve code generation and agent reasoning. MIT licensed, free to use globally, with reported strong performance on coding benchmarks (Terminal-Bench, SWE-bench, NL2Repo). Designed for efficient single-GPU deployment.

Source: HuggingFace — huggingface.co/deepreinforce-ai/Ornith-1.0-397B
396.8B
Parameters
mit
License (OSI-approved)
Unknown
Context (tokens)
94.1k
Downloads (30d)

Key facts

Objective fields from the source. Values we can't verify are shown as “Unknown” rather than guessed.

FieldValue
Developerdeepreinforce-ai
Parameters396.8B
Context windowUnknown
Licensemit — OSI-approved
Modality / tasktext-generation
Gated on HuggingFaceNo
Downloads94.1k
Likes224
Last updated2026-06-25
Sourcedeepreinforce-ai/Ornith-1.0-397B

What Ornith-1.0-397B is

Ornith-1.0-397B is a 397B-parameter mixture-of-experts (MoE) model post-trained on Qwen 3.5 and Gemma 4 base models. It employs self-improving training via RL to jointly optimize solution scaffolds and rollouts. The model supports image-text-to-text and text-generation pipelines, formatted in safetensors, and is compatible with transformers-based inference. Context length is unknown; deployment targets single-GPU inference.

Quickstart

Run Ornith-1.0-397B locally

Load the open weights with 🤗 Transformers and generate — the same model, self-hosted.

quickstart.pypython
from transformers import pipelinepipe = pipeline("text-generation", model="deepreinforce-ai/Ornith-1.0-397B")out = pipe("Explain retrieval-augmented generation in one sentence.",           max_new_tokens=128)print(out[0]["generated_text"])

Swap in vLLM or Ollama for production-grade serving. DEV.co can stand up the inference stack.

Deployment

How you'd run it

A typical self-hosted path — open weights, an inference server, your application.

DEV.co builds each layer — from GPU infrastructure to the application.

Best use cases

AI-Assisted Code Generation & Debugging

Ornith excels at Terminal-Bench 2.1 and SWE-bench tasks. Use it as a backbone for IDE plugins, code review agents, or automated bug-fixing pipelines that reason over terminal outputs and repository context.

Internal Coding Agents & Development Tools

MIT license and open-source status make it suitable for embedding in internal developer platforms, CI/CD integration, or custom LLM-based dev tools where you control the deployment and inference.

Research & Benchmarking on Agentic Coding

Ideal for teams evaluating open-source coding LLMs, studying agent scaffold discovery, or fine-tuning specialized variants for proprietary codebases without licensing friction.

Running & fine-tuning it

ESTIMATE: 397B MoE parameters at ~2 bytes per active parameter (bfloat16 with sparse activation) ≈ ~200–250GB active memory during inference. Full model ~800GB. Requires multi-GPU (8× H100 / A100 or similar) or quantized variants (int8/int4) on single high-end GPU (≥80GB). Context length unknown; confirm before production planning.

No explicit fine-tuning guidance in card. MoE models add complexity to LoRA/QLoRA (router tuning, expert selection). Feasibility depends on whether fine-tuning infra supports sparse updates. Requires review of training scripts on HuggingFace repo. Recommend consulting deepreinforce-ai community or docs before committing to fine-tuning pipeline.

When to avoid it — and what to weigh

  • Very Limited Hardware (< 80GB VRAM) — 397B parameters require substantial GPU memory, even with quantization. Not feasible on consumer-grade GPUs; requires enterprise-class hardware or cloud acceleration.
  • Production Latency-Critical Applications — MoE models introduce routing overhead. If sub-100ms inference is mandatory, smaller dense models or optimized proprietary APIs may be preferable.
  • Non-English or Domain-Specific Code Tasks — Training focused on agentic coding; coverage for non-English code, domain-specific languages (COBOL, Fortran, domain DSLs), or low-resource languages is unknown. Benchmark data is sparse.
  • Fully Offline / Air-Gapped Deployment Without Validation — While open-source and gated=false, deployment at scale in secure environments requires security audit and clearance. Not addressed in card.

License & commercial use

MIT License. Permissive open-source license allowing commercial use, modification, and distribution with attribution. No regional restrictions stated. Model weights are globally accessible (gated=false).

MIT is a permissive OSI-approved license that explicitly permits commercial use, including in proprietary products, without requiring disclosure or source-code sharing. No commercial restrictions in the license itself. However: (1) deployment at scale in production requires infrastructure and support planning; (2) model performance and liability remain your responsibility; (3) any base-model dependencies (Qwen 3.5, Gemma 4) should be audited for commercial constraints—card does not detail those licenses. Recommended: review Qwen and Gemma licensing separately before shipping.

DEV.co evaluation signals

Editorial assessment — not user reviews. Directional, with an explicit confidence level.

SignalAssessment
MaintenanceActive
DocumentationAdequate
License clarityClear
Deployment complexityHigh
DEV.co fitGood
Assessment confidenceHigh
Security considerations

No security audit, adversarial robustness, or red-teaming results disclosed. As a 397B model fine-tuned on code, it may inherit or amplify code-injection, SQL-injection, or system-prompt vulnerabilities from training data. In agentic coding scenarios (e.g., terminal execution), constrain input validation, sandbox execution, and output filtering strictly. No mention of data provenance, filtering for sensitive patterns, or training-data audits. Production deployment in security-sensitive contexts requires third-party review.

Alternatives to consider

Qwen 3.5-397B (base model)

Same parameter count, likely similar inference cost. Ornith is a post-trained variant; Qwen 3.5 base may be preferable if you want a general-purpose model and plan your own task-specific fine-tuning. No MIT license stated for Qwen 3.5; requires checking.

DeepSeek V4 Pro 1.6T

Larger, closed-source API-based alternative with strong SWE-bench results. Trade: no local control, potential vendor lock-in, but likely lower deployment complexity and commercial support.

Claude Opus 4.8 (via API)

Top SWE-bench Verified (87.6) and Pro (69.2) scores. Proprietary, managed service with commercial support. Higher cost per token, but eliminates infrastructure burden and security audit requirements for coding agent use cases.

Software development agency

Ship Ornith-1.0-397B with senior software developers

Ornith is strong for agentic coding tasks but demands significant compute. Start with a PoC on cloud infrastructure (AWS SageMaker, Anyscale, or Together AI). Audit base-model licenses and conduct a security review before production. Contact deepreinforce-ai for fine-tuning support and best practices.

Talk to DEV.co

Related open-source tools

Surfaced by semantic similarity across the DEV.co open-source index.

Related on DEV.co

Explore the category and the services that help you build with it.

Ornith-1.0-397B FAQ

Can I use Ornith-1.0-397B in a commercial product?
Yes. MIT license explicitly permits commercial use, including proprietary products. No source-code disclosure required. However, audit the base-model licenses (Qwen 3.5, Gemma 4) independently, as they may have additional constraints. Deploy with your own infrastructure and support.
What GPU hardware do I need to run this model?
Estimate: 200–250GB active VRAM for inference in bfloat16 MoE configuration. Requires multi-GPU setup (8× H100/A100) or a single high-end GPU with quantization (int8/int4). Context length is unknown; confirm that before sizing. Cloud providers (AWS, GCP) offer H100/A100 clusters on-demand.
How does Ornith compare to smaller, open-source models like CodeLlama or Granite?
Ornith-1.0-397B is orders of magnitude larger (397B vs. 7–13B). It achieves higher coding-benchmark scores (e.g., 82.4% on SWE-bench Verified vs. ~40–60% for smaller models) but requires far more compute. For latency-sensitive or resource-constrained workloads, smaller models are preferable; for agentic reasoning on complex tasks, Ornith is stronger.
Is fine-tuning supported?
Not explicitly documented in the card. MoE models complicate LoRA/QLoRA training due to sparse activation and routing. Check the deepreinforce-ai GitHub repo or contact the team for fine-tuning scripts and best practices before committing.

Software developers & web developers for hire

Adopting Ornith-1.0-397B is usually one piece of a larger software development effort. As a software development agency, DEV.co provides software development services and web development expertise — pairing senior software developers and web developers with your team to design, build, and operate open-source llms software in production.

Ready to Deploy Ornith-1.0-397B?

Ornith is strong for agentic coding tasks but demands significant compute. Start with a PoC on cloud infrastructure (AWS SageMaker, Anyscale, or Together AI). Audit base-model licenses and conduct a security review before production. Contact deepreinforce-ai for fine-tuning support and best practices.