DEV.co
Open-Source LLM · deepreinforce-ai

Ornith-1.0-397B-FP8

Ornith-1.0-397B is a 397-billion-parameter open-source LLM optimized for agentic coding tasks. It uses a mixture-of-experts (MoE) architecture built on Qwen 3.5 and employs reinforcement learning to improve code generation and reasoning. MIT-licensed, it's designed for self-hosted deployment and shows strong performance on coding benchmarks like SWE-Bench and Terminal-Bench. Best for teams needing an open, deployable coding agent without proprietary API dependencies.

Source: HuggingFace — huggingface.co/deepreinforce-ai/Ornith-1.0-397B-FP8
397B
Parameters
mit
License (OSI-approved)
Unknown
Context (tokens)
107.8k
Downloads (30d)

Key facts

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

FieldValue
Developerdeepreinforce-ai
Parameters397B
Context windowUnknown
Licensemit — OSI-approved
Modality / tasktext-generation
Gated on HuggingFaceNo
Downloads107.8k
Likes175
Last updated2026-06-25
Sourcedeepreinforce-ai/Ornith-1.0-397B-FP8

What Ornith-1.0-397B-FP8 is

Ornith-1.0-397B is a 397B-parameter MoE model post-trained on Gemma 4 and Qwen 3.5 foundations. The architecture includes a self-improving RL training framework that jointly optimizes scaffolding and solution rollouts for agentic workflows. Available in FP8 quantization (compressed-tensors format). Supports image-text-to-text and text-generation pipelines. Not gated; weights are publicly available. Context length not specified in card.

Quickstart

Run Ornith-1.0-397B-FP8 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-FP8")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

Autonomous Code Agents & AI Dev Tools

Build internal or customer-facing agents that autonomously solve coding tasks, fix bugs, or navigate repositories. Strong SWE-Bench and Terminal-Bench scores indicate reliability for end-to-end software tasks.

Self-Hosted LLM Infrastructure

Deploy as a private, on-premises coding engine to avoid API latency, data egress, or vendor lock-in. MIT license permits commercial self-hosting without licensing fees.

Fine-Tuning Base for Domain-Specific Coding

Use as a pretrained foundation for custom RL or supervised fine-tuning on proprietary codebases, internal coding standards, or niche languages.

Running & fine-tuning it

ESTIMATE: FP8 quantized 397B parameters likely require 150–250 GB VRAM for single-GPU inference (requires verification against actual model card or quantized weights). Multi-GPU setup (e.g., 8×H100/A100) recommended for throughput. Smaller variants (9B-Dense, 31B-Dense, 35B-MoE) available for constrained setups. Exact memory footprint and precision trade-offs not specified in card excerpt.

Card does not specify LoRA/QLoRA compatibility, adapter support, or fine-tuning recipes. Given the model is post-trained with RL on a base (Gemma 4 / Qwen 3.5), supervised fine-tuning or continued RL is plausible but requires explicit documentation. Recommend reviewing full model card or contacting deepreinforce-ai for guidance on custom training pipelines.

When to avoid it — and what to weigh

  • Very Limited GPU Memory — 397B parameters at FP8 still require significant VRAM. Exact quantization size unknown; rough estimate ~200GB+ for inference. Smaller variants (9B, 31B, 35B) are available if this is a blocker.
  • Natural Language or General-Purpose Reasoning as Primary Need — Model is specialized for coding. Performance on general-purpose NLU, translation, or non-technical tasks is not documented. Consider general-purpose LLMs if code is not central.
  • Guaranteed Sub-100ms Latency at Scale — Serving a 397B MoE model with consistent low latency requires significant infrastructure tuning. Verify throughput and latency SLAs with your serving stack before committing to production.
  • No In-House MLOps or GPU Infrastructure — Deploying a 397B model self-hosted demands expertise in quantization, serving frameworks (vLLM, TGI), multi-GPU setup, and monitoring. API-based services may be lower-friction.

License & commercial use

MIT License. Permissive OSI-approved license allowing use, modification, and distribution for commercial and private purposes, with attribution required. No regional restrictions; globally accessible.

MIT license clearly permits commercial use, including proprietary applications and resale, provided the license and copyright notice are retained. No gating or access restrictions. Self-hosting for production commercial services is allowed. However, verify that your intended use (e.g., SaaS deployment, fine-tuning on customer data, embedding in products) aligns with your legal/compliance requirements, particularly if you modify the model or retain proprietary training data.

DEV.co evaluation signals

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

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

No security audit, adversarial robustness analysis, or safety fine-tuning details provided in card. Code generation models can produce insecure code, dependency confusion, or injection payloads if not carefully prompted. Recommend: (1) isolate execution environments for generated code, (2) conduct internal security testing for your use case, (3) implement input validation and output sanitization, (4) monitor for unexpected behaviors. Quantization to FP8 may introduce numerical instabilities; test thoroughly before production.

Alternatives to consider

Qwen 3.5-397B (base model)

Ornith is post-trained on Qwen 3.5; if you need a general-purpose alternative without RL specialization or want to apply your own fine-tuning, the base Qwen model may be simpler and easier to customize.

DeepSeek-V4-Pro-1.6T

Larger, potentially stronger on complex reasoning (SWE-Bench Pro: 55.4 vs. Ornith's 62.2, though Ornith leads on other benchmarks). Requires more infrastructure but may offer higher ceiling for very demanding tasks.

Claude Opus 4.8 (API)

Proprietary, hosted, strong coding performance (SWE-Bench Verified: 87.6 vs. Ornith's 82.4). Avoids self-hosting complexity but introduces vendor lock-in, latency, and API costs.

Software development agency

Ship Ornith-1.0-397B-FP8 with senior software developers

Ornith-1.0-397B offers strong coding performance with full source transparency and commercial freedom. Start with infrastructure planning, GPU sizing, and a proof-of-concept on a smaller variant (9B or 35B) before scaling to 397B production workloads.

Talk to DEV.co

Related open-source tools

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

Ornith-1.0-397B-FP8 FAQ

Can I use Ornith-1.0-397B in a commercial product or SaaS application?
Yes. The MIT license permits commercial use, including embedding in SaaS products, as long as you retain the copyright and license notice. No paid licensing required. However, verify compliance with your legal/security teams if you modify the model or combine it with proprietary data.
What GPU hardware do I need to run this model?
Ornith-1.0-397B at FP8 likely requires 150–250 GB VRAM; exact footprint not specified. A single high-end GPU (e.g., H100 80GB) is insufficient. Use 4–8 GPUs (H100, A100, L40S clusters) for throughput-optimized serving. Test with your serving framework (vLLM, TGI) before production deployment. Smaller variants (9B, 31B, 35B) available if hardware is constrained.
Is the context length sufficient for long code files or multi-file reasoning?
Context length is not specified in the model card. This is critical for code-in-context (RAG) workflows. Contact deepreinforce-ai or check the full model card on HuggingFace to confirm context window before integrating into production systems.
Can I fine-tune this model on my own codebase?
Card does not document fine-tuning recipes, LoRA support, or adapter compatibility. Given the model uses RL post-training, custom RL or supervised fine-tuning is plausible but not explicitly documented. Recommend reaching out to the maintainers or reviewing GitHub repository for guidance and examples.

Work with a software development agency

From first prototype to production, DEV.co delivers software development services around tools like Ornith-1.0-397B-FP8. Our software development agency staffs experienced software developers and web developers for custom software development, web development, integrations, and ongoing support across open-source llms and beyond.

Ready to Deploy a Private Coding Agent?

Ornith-1.0-397B offers strong coding performance with full source transparency and commercial freedom. Start with infrastructure planning, GPU sizing, and a proof-of-concept on a smaller variant (9B or 35B) before scaling to 397B production workloads.