DEV.co
Open-Source LLM · deepreinforce-ai

Ornith-1.0-35B-GGUF

Ornith-1.0-35B is a 35-billion-parameter open-source language model optimized for coding tasks and agentic workflows. Built on Qwen 3.5 and Gemma 4 foundations, it uses reinforcement learning to improve both solution generation and search strategies. The model is MIT-licensed, ungated, and sized for single-GPU deployment. Benchmark results show strong performance on coding-specific tasks like SWE-Bench and Terminal-Bench compared to similarly-sized peers.

Source: HuggingFace — huggingface.co/deepreinforce-ai/Ornith-1.0-35B-GGUF
Unknown
Parameters
mit
License (OSI-approved)
Unknown
Context (tokens)
502.7k
Downloads (30d)

Key facts

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

FieldValue
Developerdeepreinforce-ai
ParametersUnknown
Context windowUnknown
Licensemit — OSI-approved
Modality / tasktext-generation
Gated on HuggingFaceNo
Downloads502.7k
Likes779
Last updated2026-06-25
Sourcedeepreinforce-ai/Ornith-1.0-35B-GGUF

What Ornith-1.0-35B-GGUF is

Ornith-1.0-35B is a 35B-parameter dense transformer model distributed in GGUF quantized format, enabling efficient inference on consumer/enterprise GPUs. The model employs a self-improving RL framework that jointly optimizes both the scaffold (search strategy) and solution rollouts for coding tasks. It is available alongside smaller (9B, 31B) and larger (397B MoE) variants in the Ornith family. Last modified June 2026; 502K downloads and 779 likes indicate moderate adoption. Context length and exact parameter count not stated in card.

Quickstart

Run Ornith-1.0-35B-GGUF 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-GGUF")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 and software engineering workflows

Ornith-1.0-35B is purpose-built for code-centric agentic tasks. Reported 75.6% on SWE-Bench Verified and 64.2% on Terminal-Bench 2.1 (Terminus-2) outperforms larger open-source baselines. Ideal for automated bug fixes, repository-scale refactoring, and test generation.

Single-GPU on-premise deployments

The 35B-Dense variant is explicitly designed for efficient single-GPU deployment. GGUF quantization support allows sub-100GB VRAM inference. Suitable for teams requiring private, self-hosted coding assistance without cloud dependencies.

Research and model improvement iteration

The self-improving RL framework and MIT license enable researchers to fine-tune, experiment with alternative scaffolds, and run custom evaluation loops. Open-source, ungated model suitable for academic and commercial research teams.

Running & fine-tuning it

ESTIMATE: 35B-Dense model in full precision (float32) requires ~140 GB VRAM; in int8 quantization ~35 GB; in int4 (GGUF) ~17–20 GB. Single-GPU deployment is feasible on A100 (80 GB), H100 (80 GB), or high-end consumer GPUs (RTX 6000, L40S). Multi-GPU inference can use tensor parallelism. Exact quantization schemes and memory overhead not detailed in card—verify with deployment tool (vLLM, TGI) before production use.

Not explicitly documented. Standard practices: LoRA fine-tuning on 35B is feasible with 24–48 GB VRAM using peft/bitsandbytes. QLoRA (4-bit + LoRA) reduces to ~8–12 GB VRAM. Self-improving RL framework may allow custom reward functions and scaffold optimization, but requires custom training infrastructure and is not a standard LoRA use case. Recommend contacting deepreinforce-ai or reviewing source for RL fine-tuning recipes.

When to avoid it — and what to weigh

  • General-purpose chat or long-form reasoning — Ornith-1.0-35B is specialized for coding tasks. General knowledge, creative writing, multilingual chat, and non-technical reasoning are not primary design targets. Consider broader models (Qwen 3.5, Gemma 4) for general-purpose use.
  • Extremely latency-sensitive applications without quantization — While GGUF quantization is available, full-precision 35B inference will require significant GPU memory and incur latency. If <100ms per-token latency is required and you cannot use quantization, evaluate smaller models or commercial APIs.
  • Deployment on commodity CPU-only systems — 35B models require GPU acceleration for practical inference. CPU-only inference is feasible but extremely slow. If CPU-only is a hard constraint, consider the 9B-Dense variant or lighter alternatives.
  • Tasks requiring certified safety or compliance guarantees — No safety training, alignment audits, or compliance certifications are documented. Use cases requiring HIPAA, PCI, or formal security/safety sign-offs require additional review and potentially custom fine-tuning.

License & commercial use

MIT License. Permissive open-source license permits commercial use, modification, and redistribution with attribution and liability disclaimer. MIT is a standard OSI-approved license with no regional or usage restrictions.

MIT license explicitly permits commercial use without restriction. Model is ungated and globally accessible. No subscription, royalties, or usage limits are stated. However, you are responsible for: (1) compliance with base model licenses (Qwen 3.5, Gemma 4) if you redistribute or fine-tune; (2) safety/compliance review for your application domain; (3) liability for model outputs. Verify base model terms and conduct legal review for high-risk use cases (healthcare, finance, autonomous systems).

DEV.co evaluation signals

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

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

No explicit security audit, adversarial testing, or red-teaming results are documented. Model is not formally tested against prompt injection, jailbreaks, or data extraction attacks. Considerations: (1) Self-hosted GGUF deployment reduces cloud data exposure; (2) RL training on coding tasks may introduce unintended behaviors—validate outputs in CI/CD before production deployment; (3) No mention of bias audits or fairness testing; (4) Input/output filtering and rate limiting are user's responsibility; (5) Review base model (Qwen, Gemma) security advisories. For sensitive applications (critical infrastructure, healthcare), conduct additional red-teaming and formal security review.

Alternatives to consider

Qwen3.5-35B (base model)

The foundation Ornith is built on. Broader general-purpose capability but lower coding-specific benchmarks (70% SWE-Bench Verified vs. Ornith's 75.6%). Better for non-coding tasks; worse for code agents.

Claude 3.5 Sonnet (commercial API)

Closed-source, API-only. Superior coding performance (real-time, production-grade). No self-hosting, privacy concerns, higher cost. Best if commercial licensing and vendor support are acceptable.

DeepSeek-Coder-33B (open-source)

Alternative 33B coding model, MIT/Apache-licensed. Performance varies by benchmark. Smaller footprint than Ornith-35B; requires direct comparison on your evaluation set. Less recent; unclear on RL-based improvements.

Software development agency

Ship Ornith-1.0-35B-GGUF with senior software developers

Ornith-1.0-35B is MIT-licensed and designed for single-GPU self-hosted deployment. Start with quantized GGUF inference using llama.cpp or vLLM, benchmark on your codebase, then fine-tune with LoRA if needed. Contact Devco to architect production deployment, security review, or custom RL integration.

Talk to DEV.co

Related open-source tools

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

Ornith-1.0-35B-GGUF FAQ

Can I use Ornith-1.0-35B commercially?
Yes. MIT license permits commercial use without royalties or restrictions. However, verify compliance with base model (Qwen, Gemma) terms and conduct safety/compliance review for your application domain (healthcare, finance, autonomous systems may require additional vetting).
What GPU do I need to run this model?
Minimum: ~24–35 GB VRAM for quantized inference (int4/int8 GGUF). Recommended: A100 (80 GB), H100, or equivalent for full precision or multi-user serving. RTX 4090 (24 GB) can run int4 GGUF with tight memory management. Use llama.cpp or vLLM with quantization for consumer GPUs.
Can I fine-tune Ornith-1.0-35B?
Standard LoRA/QLoRA fine-tuning is supported but not explicitly documented in the card. Custom RL scaffold optimization is possible but requires significant infrastructure. Contact deepreinforce-ai or review the GitHub repo for training recipes.
How does Ornith-1.0-35B compare to Claude or GPT-4?
On coding benchmarks (SWE-Bench, Terminal-Bench), Ornith-35B outperforms similarly-sized open models and approaches larger commercial models in some tasks. However, closed-source APIs (Claude, GPT-4) have ongoing updates, superior general knowledge, and production SLAs. Ornith is best for self-hosted, cost-controlled coding automation; Claude/GPT-4 for maximum capability and vendor support.

Work with a software development agency

DEV.co is a software development agency delivering custom software development services to companies building on open source. Our software developers and web developers design, integrate, and ship production systems — spanning web development, APIs, AI, data, and cloud. If Ornith-1.0-35B-GGUF is part of your open-source llms roadmap, our team can implement, customize, migrate, and maintain it.

Ready to deploy a private coding agent?

Ornith-1.0-35B is MIT-licensed and designed for single-GPU self-hosted deployment. Start with quantized GGUF inference using llama.cpp or vLLM, benchmark on your codebase, then fine-tune with LoRA if needed. Contact Devco to architect production deployment, security review, or custom RL integration.