DEV.co
Open-Source LLM · SC117

Ornith-1.0-35B-MTP-APEX-GGUF

Ornith-1.0-35B-MTP-APEX-GGUF is a 35B-parameter mixture-of-experts LLM optimized for agentic coding tasks. It is derived from Qwen3.5 and trained with reinforcement learning for code generation and reasoning. The GGUF quantization makes it suitable for local/on-premise deployment. MIT license permits commercial use without restrictions. Not gated.

Source: HuggingFace — huggingface.co/SC117/Ornith-1.0-35B-MTP-APEX-GGUF
Unknown
Parameters
mit
License (OSI-approved)
Unknown
Context (tokens)
47.5k
Downloads (30d)

Key facts

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

FieldValue
DeveloperSC117
ParametersUnknown
Context windowUnknown
Licensemit — OSI-approved
Modality / tasktext-generation
Gated on HuggingFaceNo
Downloads47.5k
Likes45
Last updated2026-07-02
SourceSC117/Ornith-1.0-35B-MTP-APEX-GGUF

What Ornith-1.0-35B-MTP-APEX-GGUF is

Qwen3.5-based MoE model with 256 routed experts (3B active per token), 40 transformer layers + 1 MTP layer, 262k context window, and integrated vision projector (mmproj) for multimodal capabilities. Distributed as APEX-quantized GGUFs optimized for llama.cpp. MTP layer sourced from Qwen3.5-35B-A3B. Benchmarked on Terminal-Bench 2.1, SWE-Bench Verified/Pro/Multilingual, NL2Repo, and OpenClaw.

Quickstart

Run Ornith-1.0-35B-MTP-APEX-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="SC117/Ornith-1.0-35B-MTP-APEX-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

On-Premise Code Generation & Debugging

GGUF quantization and llama.cpp compatibility enable deployment on consumer/enterprise hardware without cloud dependency. Ideal for teams requiring code IP to remain internal. MoE architecture keeps memory footprint reasonable (3B active tokens) despite 35B total parameters.

SWE-Bench Workflow Automation

Model demonstrates strong performance on SWE-Bench Verified/Pro/Multilingual. Suitable for automated software engineering tasks: pull request analysis, bug localization, multi-file code changes, and repository context understanding via extended context window.

Multimodal Code Assistance

Integrated mmproj vision projector allows processing screenshots, diagrams, and UI images alongside code. Useful for visual debugging tools, documentation extraction from images, and diagram-to-code workflows.

Running & fine-tuning it

ESTIMATE: APEX-I-Compact quantization ~15.85 GB (per model card). Full precision likely 140+ GB. For local inference: minimum 16 GB GPU VRAM (RTX 4060 Ti / A6000 range) or 32+ GB CPU RAM. Benchmarks reference RTX 5070 Ti. Throughput and latency scale with hardware; consumer GPUs will see ~1–10 tokens/sec depending on quantization and batch size. Verify VRAM headroom by testing on target hardware.

Not documented in card. Model is post-trained on RL; no mention of LoRA/QLoRA compatibility or fine-tuning recommendations. If fine-tuning is required, consult base model (Ornith-1.0-35B) or Qwen3.5 documentation. Quantized GGUFs are typically not fine-tuned; dequantize or obtain unquantized weights if adaptation is needed.

When to avoid it — and what to weigh

  • Latency-Critical Real-Time Systems — GGUF inference on consumer GPUs/CPUs will exhibit higher latency than commercial API endpoints. 35B parameters (even with MoE sparsity) demand significant compute. Not suitable for sub-100ms response SLAs.
  • Production Deployments Without Infrastructure — While GGUF simplifies local serving, production use still requires GPU hardware (RTX 5070 Ti+ implied from benchmarks), vLLM/TGI setup, monitoring, and failover planning. Not a turnkey solution.
  • Need for Domain Specialization via Fine-Tuning — No LoRA/QLoRA guidance documented. Model is post-trained and likely not optimized for further fine-tuning. If your domain requires significant adaptation, alternatives with explicit fine-tuning support may be preferable.
  • Minimal Hardware Constraints — APEX-I-Compact is 15.85 GB; even quantized, this exceeds mobile, edge, or very-low-memory deployments. Requires dedicated GPU VRAM or substantial CPU memory.

License & commercial use

MIT License. Permissive, open-source. Allows commercial use, modification, and distribution with attribution. No copyleft or proprietary restrictions.

MIT is a permissive OSI-approved license. Commercial use is explicitly permitted without restrictions. No dual licensing, no special commercial terms, no gating. You may deploy, modify, and sell products built on this model. Attribution required. No liability warranty—typical for open-source.

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

GGUF quantization reduces attack surface vs. full-precision weights. Local deployment (llama.cpp) avoids cloud data transmission. No security audit mentioned. Standard LLM risks apply: prompt injection, model extraction (if served over network), and resource exhaustion (DoS via large context requests). Validate input sizes and implement rate-limiting if exposed. No built-in content filtering noted.

Alternatives to consider

Qwen3.5 (base model, unquantized)

Original Qwen3.5 MoE backbone. Offers more flexibility for fine-tuning and integration with ecosystem tools. Less optimized for on-prem GGUF inference; requires cloud or high-end hardware.

DeepSeek-Coder v2 (236B, open-source)

Larger open-source coding model with strong SWE-Bench performance. More parameters but also higher resource demand. Different licensing and deployment profile.

Llama 3.1 (70B/8B variants, Meta)

Broader generalist model with extensive community support and documentation. Lower coding specialization but wider ecosystem integration (vLLM, TGI, Ollama). MIT license.

Software development agency

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

Evaluate this open-source agentic coding model on your infrastructure. GGUF quantization and llama.cpp support enable zero-cloud-dependency deployment. MIT license permits commercial use. Start with APEX-I-Compact (15.85 GB) and scale with our guidance on hardware, serving, and integration.

Talk to DEV.co

Related open-source tools

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

Ornith-1.0-35B-MTP-APEX-GGUF FAQ

Can I use this model commercially?
Yes. MIT license permits unrestricted commercial use, modification, and distribution. Attribution required. No licensing fees or special commercial agreement needed. Ideal for SaaS, proprietary products, and enterprise deployment.
How much GPU memory do I need to run Ornith-1.0-35B-MTP-APEX locally?
APEX-I-Compact is ~15.85 GB. Recommend 24+ GB GPU VRAM (e.g., RTX 4090, A100 40GB) for comfortable inference with batch processing. Smaller quantizations or CPU offloading possible on 16 GB but expect slower throughput (1–5 tokens/sec). Test on your target hardware to confirm.
Does this model support fine-tuning?
Not explicitly documented. Model is post-trained and distributed as quantized GGUF. Fine-tuning is not recommended for quantized weights. If you need domain-specific adaptation, request unquantized weights from SC117/DeepReinforce AI or consult Qwen3.5 fine-tuning guides.
What is MTP and how does it affect performance?
MTP (likely 'Multi-Token Prediction' or similar) is an additional layer sourced from Qwen3.5-35B-A3B architecture, integrated for improved reasoning/generation. Benchmark data shows minimal performance delta (93.5 'Thinking' vs. 93.2 'No Thinking' on max benchmark). MTP may improve sample efficiency; exact mechanism not detailed in card.

Custom software development services

Need help beyond evaluating Ornith-1.0-35B-MTP-APEX-GGUF? DEV.co is a software development agency offering software development services and web development for teams of every size. Our software developers and web developers build custom software, web applications, APIs, and open-source llms integrations — and maintain them long-term.

Deploy Ornith-1.0-35B Locally for Secure Code Workflows

Evaluate this open-source agentic coding model on your infrastructure. GGUF quantization and llama.cpp support enable zero-cloud-dependency deployment. MIT license permits commercial use. Start with APEX-I-Compact (15.85 GB) and scale with our guidance on hardware, serving, and integration.