mirxa2/Ornith-1.0-35B-Ultimate-BF16

⸻

license: mit license_link: https://huggingface.co/deepreinforce-ai/Ornith-1.0-35B/blob/main/LICENSE base_model: deepreinforce-ai/Ornith-1.0-35B library_name: transformers pipeline_tag: text-generation author: Mirxa27 tags:

• qwen3_5_moe
• gated-deltanet
• hybrid
• moe
• mixture-of-experts
• reasoning
• thinking
• coding
• agentic
• swe-bench
• terminal-bench
• tool-calling
• vision
• multimodal
• image-text-to-text
• norm-preserving-biprojection
• vllm
• dgx-spark
• gb10
• bfloat16
• conversational
• 35b

⸻

Ornith-1.0-35B-Mirxa27-AEON-BF16

An optimized build of deepreinforce-ai/Ornith-1.0-35B, DeepReinforce’s advanced agentic-coding Mixture-of-Experts model, prepared and released by Mirxa27 with a focus on capability preservation, long-context reasoning, coding workflows, and agentic tool-use performance.

Lineage note: the upstream card states that the family is post-trained on Gemma 4 and Qwen 3.5. Weight-correlation testing indicates that the 35B-MoE member is initialized from Qwen/Qwen3.6-35B-A3B. Architecture: qwen3_5_moe — 40 layers, including 30 GatedDeltaNet linear-attention layers and 10 full-attention layers, 256 routed experts plus 1 shared expert, A3B routing, vision tower, and 256K context support.

Model Overview

This release is designed for demanding reasoning, coding, tool-calling, and multimodal workflows while maintaining the strengths of the original Ornith-1.0-35B base model.

Key areas of focus:

1. Agentic coding
   • Strong software engineering performance.
   • Suitable for code generation, debugging, refactoring, terminal workflows, and tool-using agents.
2. Long-context reasoning
   • Supports extended context deployments up to 256K tokens when infrastructure allows.
   • Useful for large codebases, document analysis, multi-step planning, and agent memory workflows.
3. Tool calling
   • Compatible with vLLM tool-calling workflows.
   • Designed for agentic systems where structured actions and external tools are required.
4. Multimodal capability
   • Vision support is inherited from the base model.
   • Suitable for image-text and multimodal assistant pipelines when deployed with compatible infrastructure.
5. Capability preservation
   • This build is prepared to keep the original model’s core reasoning, coding, and multimodal capabilities stable.

Technical Notes

This model keeps the core architecture and major capability profile of the upstream Ornith-1.0-35B release.

Architecture highlights:

• qwen3_5_moe
• 35B Mixture-of-Experts family
• 40 total layers
• 30 GatedDeltaNet linear-attention layers
• 10 full-attention layers
• 256 routed experts plus 1 shared expert
• Vision tower support
• Long-context support up to 256K tokens
• BF16 precision release

Validation Summary

Metric Original Ornith-1.0-35B This model Agentic/coding pass@1, 18-task probe 0.833 0.833 First-token KL vs base — ~0.0014 Long-generation coherence — Stable Vision support Supported Supported Tool-calling support Supported Supported

The model is intended to preserve the original Ornith family’s core competency in agentic coding, reasoning, multimodal use, and long-context workflows.

Reference upstream scores for Ornith include:

• Terminal-Bench 2.1: 64.2
• SWE-bench Verified: 75.6

Quickstart with vLLM

vllm serve Mirxa27/Ornith-1.0-35B-Mirxa27-AEON-BF16
--served-model-name ornith
--max-model-len 262144
--gpu-memory-utilization 0.85
--max-num-batched-tokens 16384
--mamba-cache-dtype float32
--reasoning-parser qwen3
--enable-auto-tool-choice
--tool-call-parser qwen3_coder
--limit-mm-per-prompt '{"image":4,"video":2}'
--mm-encoder-tp-mode data
--attention-backend flash_attn
--enable-chunked-prefill
--enable-prefix-caching
--trust-remote-code

--served-model-name can be used to provide one or more aliases for easier integration with existing clients or OpenAI-compatible endpoints.

Recommended sampling:

temperature: 0.6 top_p: 0.95 top_k: 20

For reasoning workflows, the model may produce structured reasoning-style output depending on the serving stack, template, and client configuration.

Deployment Notes

For best results:

• Use vLLM with compatible GPU infrastructure.
• Enable prefix caching for repeated agentic or chat workloads.
• Use chunked prefill for long-context workloads.
• Keep KV cache and multimodal settings aligned with your hardware limits.
• Validate context length before production deployment, especially when targeting 128K to 256K token windows.

For vision-enabled deployments, ensure the serving environment supports the required multimodal pipeline and memory allocation.

Variants & Quantization

• BF16 — full precision release. Approximate storage requirement: around 66 GB.
• Low-precision variants may be prepared separately depending on deployment needs and available GPU hardware.

For high-performance production usage, test the exact quantization format on your target hardware before deployment. Some low-precision formats may reduce quality or coherence depending on kernels, GPU generation, and serving configuration.

Responsible Use

This model is a powerful general-purpose AI system intended for legitimate research, development, automation, coding, reasoning, and multimodal applications.

Users are responsible for ensuring that their use of the model complies with applicable laws, platform rules, organizational policies, and safety requirements.

Recommended production safeguards include:

1. Input validation
   • Validate user inputs before sending them to the model.
2. Output review
   • Review or filter outputs when used in sensitive, public-facing, or automated environments.
3. Access controls
   • Limit access to authorized users and trusted systems.
4. Audit logging
   • Keep logs for debugging, monitoring, and abuse prevention where legally permitted.
5. Human oversight
   • Use human review for high-impact workflows, financial decisions, legal content, medical content, security-sensitive tasks, or automated system actions.
6. Tool-use protection
   • When the model is connected to external tools, APIs, terminals, browsers, databases, or production systems, apply strict permissions and confirmation steps.
7. Testing before deployment
   • Evaluate the model carefully on your own workload before using it in production.

This model is provided as a tool. The user or deploying organization remains responsible for how it is configured, accessed, monitored, and used.

Disclaimer

This model is provided “AS IS”, without warranty of any kind, express or implied, including but not limited to warranties of merchantability, fitness for a particular purpose, non-infringement, factual accuracy, safety, or legal compliance.

The authors, contributors, publishers, and hosting platforms do not assume liability for outputs or downstream usage. Users are responsible for validating outputs and ensuring lawful, ethical, and appropriate deployment.

License

This release inherits the MIT license from the base model.

• Base model: deepreinforce-ai/Ornith-1.0-35B
• License: MIT

Provenance & Credits

• Base: deepreinforce-ai/Ornith-1.0-35B
• Build / release: Mirxa27
• Architecture family: qwen3_5_moe
• Methods referenced: norm-preserving biprojection, expert-level tuning, SSM stability repair, long-context validation
• Serving stack: Transformers / vLLM compatible

License: MIT, inherited from the base model.