Ornith 1.5 35B-A3B Benchmarks: MoE Model Outperforms Qwen3.6 in Coding and Agentic Tasks

ArticleLuis Chavez-MattosNews

Deep Reinforce has introduced Ornith 1.5 35B-A3B, a mixture-of-experts model activating roughly 3 billion parameters per token out of 35 billion. According to published benchmarks, it surpasses direct competitors like Qwen3.6-35B across coding and agentic tasks, while closing the gap on much larger models such as Qwen3.5-397B.

Key points

  1. Mixture-of-experts architecture with 35B total parameters and approximately 3B active parameters per token.
  2. Trained using a joint self-improvement loop with GRPO that optimises task generation, scaffolds, and solution rollouts simultaneously.
  3. Scores 79 on SWE-bench Verified and 22 on DeepSWE, where competitors like Qwen3.6-35B and Gemma 4-31B scored 0.
  4. Trails larger models on tool-assisted pure reasoning tests like HLE (33.4 vs 48.3 for Qwen3.5-397B).
  5. Runs locally on a single Nvidia A100 GPU (80GB VRAM) under vLLM, using roughly 74GB memory including KV cache.

Why it matters

Teams evaluating local or self-hosted models for software engineering and agentic workflows benefit from low inference overhead on single-GPU hardware while achieving benchmark performance comparable to much larger dense architectures.

Analysis

The stark margin on DeepSWE highlights the effectiveness of self-generating training curricula using reinforcement learning. In real-world community tests, the model independently planned, coded, and verified multi-container features without prescriptive guidance. However, its lower performance on HLE with tools confirms that larger parameter scales remain superior for broad, pure reasoning benchmarks.

Mentioned links

Found in

Summary generated by AI (Gemini 3.8 Flash, 27 September 2026). It may contain errors – the original source is authoritative.

Related entries

Similar in content, found by the AI search.

  • Link:ornith-ai

    Ornith-1.5-397B: Open-Source MoE Model Focused on Coding and Reasoning

    ornith-ai has released Ornith-1.5-397B, an open-source multimodal mixture-of-experts model under the MIT licence. With 397 billion parameters and an end-to-end self-improvement loop, it delivers coding and reasoning benchmark performance comparable to top proprietary models.

    AI & LLMs· Resource

  • Link:deep-reinforce.com

    Ornith-1.0: Self-Scaffolding Open-Source Models for Agentic Coding

    Ornith-1.0 is an open-source model family tailored for agentic coding, spanning parameter sizes from 9B to 397B. Built on Gemma 4 and Qwen 3.5, the architecture trains models to jointly generate both code solutions and the task-specific scaffolds directing them.

    AI & LLMs· Announcement

  • Repository:ornith-ai/Ornith-1

    Ornith: Self-Improving Open-Source Models for Agentic Workflows

    The Ornith-1 repository provides an open-source model family designed for agentic tasks using autonomous self-improvement. Spanning Ornith-1.0 and Ornith-1.5, the project implements self-scaffolding and automated generation of training rollouts for reinforcement learning.

    2kstars

    AI & LLMs· Tool

  • Link:Kwaipilot

    KAT-Coder-V2.5-Dev: Open-Weight MoE Model for Coding Agents

    Kwaipilot has released KAT-Coder-V2.5-Dev, an open-weight Mixture of Experts model based on Qwen3.6-35B-A3B. Featuring 35 billion total and 3 billion activated parameters, it has been specifically post-trained for agentic software engineering tasks.

    AI & LLMs· Tool

  • Video

    Video:xCreate

    Hands-On Test: Ornith 397B Evaluated Against Qwen 3.5 and GLM 5.2

    xCreate reviews Ornith 1.0 (397B), an open-source fine-tune based on Qwen 3.5 claiming to outperform models like GLM 5.2 and Claude. The video evaluates its real-world coding, vision, and logic capabilities on a Mac Studio.

    AI & LLMs· Demo

  • Article:Chris Alexiuk and Chintan Patel

    NVIDIA Nemotron 3.5 Lightning: Compact MoE Model for Agentic Execution

    NVIDIA has announced Nemotron 3.5 Lightning, an open 30B mixture-of-experts (MoE) model featuring 3B active parameters. It is tailored for the high-volume execution layer of long-running AI agents, targeting routine tasks such as tool calling and result validation with low latency.

    AI & LLMs· Announcement

Back to the knowledge base

Let's talk about your project

Get in touch

Locations

  • Mattersburg
    Johann Nepomuk Bergerstraße 7/2/14
    7210 Mattersburg, Austria
  • Vienna
    Ungargasse 64-66/3/404
    1030 Wien, Austria

Parts of this content were created with the assistance of AI.