DEV.co
Open-Source LLM · deepreinforce-ai

Ornith-1.0-35B-FP8

Ornith-1.0-35B is a 35-billion-parameter open-source LLM optimized for agentic coding tasks. It is post-trained on Qwen 3.5 and uses a mixture-of-experts (MoE) architecture. The model is MIT-licensed, ungated, and designed for single-GPU deployment. It reports strong performance on coding benchmarks (SWE-Bench, Terminal-Bench, NL2Repo) relative to comparable-sized open models.

Source: HuggingFace — huggingface.co/deepreinforce-ai/Ornith-1.0-35B-FP8
35.1B
Parameters
mit
License (OSI-approved)
Unknown
Context (tokens)
119.8k
Downloads (30d)

Key facts

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

FieldValue
Developerdeepreinforce-ai
Parameters35.1B
Context windowUnknown
Licensemit — OSI-approved
Modality / tasktext-generation
Gated on HuggingFaceNo
Downloads119.8k
Likes65
Last updated2026-06-26
Sourcedeepreinforce-ai/Ornith-1.0-35B-FP8

What Ornith-1.0-35B-FP8 is

Ornith-1.0-35B is a 35B-parameter MoE model built on Qwen 3.5. It employs reinforcement learning to jointly optimize solution scaffolds and rollouts for code generation. The model is quantized to FP8 precision, available in SafeTensors format, and tagged as endpoints-compatible. Last modified June 2026. Downloads: 119,830; Likes: 65. Context length and inference latency are not documented.

Quickstart

Run Ornith-1.0-35B-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-1.0-35B-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

Software Engineering Automation

Strong performance on SWE-Bench (75.6% Verified, 50.4% Pro) and Terminal-Bench (64.2%) makes it suitable for automated bug fixing, repository navigation, and multi-step code generation tasks where agents must plan and execute scaffold-driven workflows.

Single-GPU Coding Workloads

35B-parameter MoE with FP8 quantization enables efficient deployment on consumer/mid-tier GPUs for real-time code completion and agentic coding tasks without requiring large distributed inference clusters.

Open-Source AI Agent Development

MIT license and ungated model allow commercial and research integration into proprietary coding agents, RAG pipelines, or IDE plugins without licensing friction.

Running & fine-tuning it

ESTIMATE (requires verification): FP8 quantization of 35B parameters ≈ ~70–80 GB raw model size. Single-GPU deployment suggested by model card implies targeting H100/A100 or equivalent (~80GB VRAM). Running on smaller GPUs (RTX 4090, 48GB) would likely require further quantization (INT4/INT8 via LoRA or distillation) or offloading. No official specs provided; benchmark actual inference on target hardware.

Model card does not explicitly discuss fine-tuning, LoRA, or QLoRA feasibility. Given the MoE architecture and RL-based post-training, efficient adaptation is plausible but not confirmed. LoRA on a 35B MoE model is more complex than dense models; QLoRA may be viable on 24–48GB GPUs. Recommend consulting deepreinforce-ai documentation or running proof-of-concept before committing to large-scale fine-tuning.

When to avoid it — and what to weigh

  • Requirement for Very Long Context (>32K tokens) — Context length is not documented. If your use case requires extended reasoning windows or large codebase analysis, context limitations are unknown and must be tested.
  • Need for General-Purpose Conversation — Model is specialized for coding and agentic tasks. Performance on general QA, creative writing, or non-technical domains is not benchmarked and likely inferior to generalist models.
  • Real-Time, Sub-100ms Latency Requirement — 35B parameter count and MoE architecture trade computational cost for quality. Inference speed and serving overhead are not disclosed; verify latency expectations against measured throughput.
  • Requirement for Proven Production Stability at Scale — Recent release (June 2026), moderate download count (119k), and no disclosed production deployment references. Suitable for early adoption or controlled pilots; not yet battle-tested in large production systems.

License & commercial use

MIT license. Permissive, OSI-compliant open-source license allowing commercial use, modification, and redistribution with attribution.

MIT license explicitly permits commercial use. No gating, no regional restrictions, no additional terms stated. Model weights and derivatives can be integrated into proprietary products. However: (1) verify terms of underlying base models (Qwen 3.5, Gemma 4) if derivatives are used; (2) consult legal counsel before deploying in regulated domains (healthcare, finance, etc.) due to lack of documented safety testing and compliance vetting.

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 explicit security audit, adversarial testing, or safety training reported. Model is trained on public code datasets (SWE-Bench, Terminal-Bench); risk of embedding known vulnerabilities or malicious patterns in training data is not assessed. FP8 quantization may introduce minor numerical drift; impact on model behavior is untested. Use in security-critical or safety-sensitive contexts requires independent red-teaming and validation.

Alternatives to consider

Qwen3.5-35B (base, non-specialized)

Same parameter count and architecture, but not fine-tuned for coding. Offers flexibility for general tasks but lacks agentic optimization and benchmark parity on code tasks.

DeepSeek-Coder-33B or similar

Comparable size, also optimized for code. May offer different trade-offs in inference speed, licensing terms, or benchmark coverage; requires comparison testing.

Gemma 4 31B (base)

Slightly smaller, also recent. Offers different licensing (to verify) and architectural choices; benchmark results shown in card suggest Ornith-1.0-35B outperforms on most metrics but Gemma 4 may better suit other workflows.

Software development agency

Ship Ornith-1.0-35B-FP8 with senior software developers

Ornith-1.0-35B offers MIT-licensed, agentic coding capability with proven benchmarks on real-world tasks. Start with a single-GPU pilot, test context length and latency on your hardware, and validate security posture before production rollout. Contact Devco for architecture guidance or custom fine-tuning.

Talk to DEV.co

Related open-source tools

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

Ornith-1.0-35B-FP8 FAQ

Can I use Ornith-1.0-35B in a commercial product?
Yes. MIT license permits commercial use without royalties or permission. However, verify compliance with underlying base models' terms (Qwen 3.5, Gemma 4) and conduct security/compliance review before deploying in regulated sectors.
What GPU do I need to run this model?
Model card suggests single-GPU deployment with FP8 quantization. Estimated VRAM: 70–80 GB for full model, or ~24–48 GB if further quantized. H100/A100 recommended; smaller GPUs require additional quantization or offloading. Benchmark on your hardware before production use.
How do I fine-tune Ornith-1.0-35B for my own coding dataset?
Not documented in model card. LoRA/QLoRA feasibility is plausible but unconfirmed for MoE architecture. Recommend: (1) check deepreinforce-ai GitHub or documentation; (2) run small proof-of-concept; (3) consider using base Qwen 3.5 if fine-tuning support is critical.
What is the model's context length?
Unknown. Not stated in model card. Likely inherits from Qwen 3.5 base (~32K or similar), but verify via inference test or official documentation before assuming for long-context tasks.

Software developers & web developers for hire

From first prototype to production, DEV.co delivers software development services around tools like Ornith-1.0-35B-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 Coding LLM?

Ornith-1.0-35B offers MIT-licensed, agentic coding capability with proven benchmarks on real-world tasks. Start with a single-GPU pilot, test context length and latency on your hardware, and validate security posture before production rollout. Contact Devco for architecture guidance or custom fine-tuning.