DEV.co
Open-Source LLM · deepreinforce-ai

Ornith-1.0-35B

Ornith-1.0-35B is a 35B-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. MIT-licensed, it's designed for single-GPU deployment and shows strong benchmark performance on code-related tasks like SWE-Bench and Terminal-Bench.

Source: HuggingFace — huggingface.co/deepreinforce-ai/Ornith-1.0-35B
664944
Parameters
mit
License (OSI-approved)
Unknown
Context (tokens)
280.2k
Downloads (30d)

Key facts

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

FieldValue
Developerdeepreinforce-ai
Parameters664944
Context windowUnknown
Licensemit — OSI-approved
Modality / tasktext-generation
Gated on HuggingFaceNo
Downloads280.2k
Likes357
Last updated2026-06-25
Sourcedeepreinforce-ai/Ornith-1.0-35B

What Ornith-1.0-35B is

Ornith-1.0-35B is a MoE-based language model with 664,944 total parameters (exact active parameter count unknown). Architecture is post-trained on Qwen 3.5 MoE base. Training uses a self-improving RL framework that jointly optimizes scaffolding and solution rollouts for code tasks. Supports text-generation and image-text-to-text pipelines. Context length unknown. Distributed as safetensors format, compatible with transformers ecosystem.

Quickstart

Run Ornith-1.0-35B 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-1.0-35B")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 Agent Development

Build agents that autonomously resolve GitHub issues or complete code tasks. Model shows 75.6% on SWE-Bench Verified and 50.4% on SWE-Bench Pro, making it suitable for agentic scaffolding in software engineering workflows.

Terminal/CLI Task Automation

Deploy as a command-planning agent for terminal automation. Terminal-Bench scores (64.2% Terminus-2, 62.8% Claude Code) indicate strong capability for planning and executing shell-like tasks.

Self-Hosted Code Completion in Private Environments

MIT license and non-gated model enable private deployment for code completion on sensitive codebases without external API calls. Single-GPU design reduces infrastructure overhead.

Running & fine-tuning it

ESTIMATE: ~70–140 GB VRAM for fp32 inference depending on MoE sparsity activation. Realistically, bf16/fp16 quantization (~35–70 GB) or GPTQ/AWQ quantization (~18–25 GB) typical for single-GPU deployment. Exact sparsity and activation patterns not documented—requires benchmarking. Model card claims single-GPU deployment feasibility but does not specify GPU tier or memory.

Not documented. LoRA/QLoRA feasibility unknown. Given MoE architecture and RL-based training approach, standard fine-tuning guidance insufficient. Requires custom exploration or guidance from deepreinforce-ai.

When to avoid it — and what to weigh

  • Long-Context or Document-Heavy Tasks — Context length is unknown and not documented. If you need extended context (>8k tokens), verify feasibility before relying on this model.
  • General-Purpose Language Understanding at Scale — Model is specialized for coding and agentic tasks. General NLP benchmarks (MMLU, etc.) are not provided. Use general-purpose alternatives if broad language capability is primary requirement.
  • Real-Time, Sub-Second Latency Requirements — 35B MoE parameters require significant compute. Inference latency not documented. For ultra-low latency, benchmark against smaller models or quantized versions.
  • Regulated Industries Without Security Review — No security audit or compliance documentation provided. Pre-deployment security review required for healthcare, finance, or other regulated sectors.

License & commercial use

MIT License. Permissive OSI-approved license allowing modification, distribution, and commercial use without restriction, provided original license is retained.

MIT license explicitly permits commercial use. No gating, regional restrictions, or additional commercial terms stated. Can be deployed in production and monetized. However, verify that deepreinforce-ai has not published separate commercial terms or usage policies outside the license file.

DEV.co evaluation signals

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

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

No security audit, adversarial robustness testing, or safety evaluation documented. Code generation models inherit risks: generated code may contain vulnerabilities or introduce supply-chain issues if deployed without review. Recommend: (1) validate generated code before execution; (2) run in sandboxed environments; (3) audit training data provenance (not disclosed); (4) test for prompt injection and jailbreaks in agentic context.

Alternatives to consider

Qwen 3.5-35B (base model)

Parent model; likely more stable but not optimized for agentic coding. Compare benchmarks directly if general-purpose performance needed.

Gemma 4-31B

Similar size, different architecture. Benchmarks show variable performance (lower on SWE-Bench Verified at 52%, but open-source alternative).

Claude 3.5 Sonnet (proprietary API)

Closes performance gap on coding benchmarks but requires external API, vendor lock-in, and per-token pricing. Consider if latency/cost trade-offs favor managed service.

Software development agency

Ship Ornith-1.0-35B with senior software developers

Benchmark Ornith-1.0-35B on your code tasks and infrastructure. Contact our team to discuss integration into private LLM or custom AI application workflows.

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-1.0-35B FAQ

Can I use Ornith-1.0-35B commercially?
Yes. The MIT license permits commercial use without restriction. No gating or licensing fees apply. Verify deepreinforce-ai has not published additional commercial terms outside the license.
What GPU should I use to run this model?
Model card claims single-GPU deployment but does not specify GPU tier or VRAM. Estimate 35–70 GB for fp16/bf16 quantization. A40, A100 (40GB+), or H100 likely suitable. Benchmark with your target hardware before production.
What is the context length?
Unknown. Not documented in the model card. Check HuggingFace model config or contact deepreinforce-ai. Critical for tasks requiring long documents.
Can I fine-tune this model?
Not clearly documented. MoE architecture and RL training approach differ from standard fine-tuning. LoRA/QLoRA feasibility unknown. Recommend testing or consulting deepreinforce-ai documentation.

Software development & web development with DEV.co

Adopting Ornith-1.0-35B 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 for Autonomous Coding?

Benchmark Ornith-1.0-35B on your code tasks and infrastructure. Contact our team to discuss integration into private LLM or custom AI application workflows.