Ornith 1.5 9B โ€” BuildEng V9 (GGUF)

Engineering-focused fine-tune by Irfan Uruchi.

Base model: ornith-ai/Ornith-1.5-9B
Published formats: Q4_K_M, Q5_K_M, Q6_K
Intended interface: Text in, text out

BuildEng V9 adapts Ornith 1.5 9B for technical discussions of building systems and construction engineering: structural load paths, reinforced-concrete distress, construction changes, temporary works, engineering assessment, and the identification of missing assumptions or required checks.

This is a technical discussion and drafting assistant, not a structural-analysis solver, a code-compliance checker, or an independently validated engineering decision system.

Available quantizations

File Format Approximate download size When to choose it
Ornith-1.5-9B-BuildEng-V9-Q4_K_M.gguf Q4_K_M 5.68 GB Smallest download; recommended starting point for limited memory.
Ornith-1.5-9B-BuildEng-V9-Q5_K_M.gguf Q5_K_M 6.49 GB A middle ground between weight precision and storage.
Ornith-1.5-9B-BuildEng-V9-Q6_K.gguf Q6_K 7.36 GB Highest weight precision among these GGUF releases.

Sizes are approximate decimal gigabytes. No comparative GGUF quality, throughput, latency, or peak-memory benchmarks have been established. Required runtime memory exceeds the model file size and depends on the backend, context length, and KV cache.

Usage

Use a recent llama.cpp build with support for this Qwen3.5-derived architecture. The GGUF files can also be loaded in LM Studio or Jan when their bundled inference runtimes support the architecture; these applications have not been independently tested with this release.

Download Q4_K_M with the Hugging Face CLI:

hf download Irfanuruchi/Ornith-1.5-9B-BuildEng-V9-GGUF \
  Ornith-1.5-9B-BuildEng-V9-Q4_K_M.gguf \
  --local-dir ./buildeng-v9

Run a single-turn text-generation check:

llama-cli \
  -m ./buildeng-v9/Ornith-1.5-9B-BuildEng-V9-Q4_K_M.gguf \
  --jinja \
  --reasoning off \
  --conversation \
  --single-turn \
  -p "Explain the load path of a transfer beam and identify the checks required before changing its support conditions."

--reasoning off requests non-thinking responses; the exact behavior depends on the model chat template and llama.cpp version. Check llama-cli --help if your build does not recognize an option.

These GGUF files are intended for text generation. Image, audio, and other multimodal inputs have not been validated for this release.

Training and conversion provenance

BuildEng V9 is derived from Ornith 1.5 9B, using domain-focused fine-tuning associated with the BuildEng dataset.

The GGUF conversion followed these steps:

  1. Adapter fusion: Fuse the BuildEng V9 LoRA adapter into an Ornith 1.5 9B MLX 8-bit base.
  2. Dequantization: Export the fused quantized representation as floating-point tensors. Dequantization does not recover the original BF16 precision.
  3. Tensor canonicalization: Reverse MLX-specific Qwen3.5 normalization offsets and restore the expected convolution tensor layout for conversion.
  4. GGUF export: Convert the canonicalized checkpoint to F16 GGUF using llama.cpp tooling.
  5. Quantization: Create Q4_K_M, Q5_K_M, and Q6_K GGUF distributions.

The GGUF files are not direct repackagings of the published MLX 4-bit checkpoint. They use the V9 adapter fused into an MLX 8-bit base, and numerical equivalence between these release paths has not been established.

Related MLX releases

Validation status

Completed:

  • Functional text-generation smoke tests for Q4_K_M, Q5_K_M, and Q6_K in the release workflow.
  • Hugging Face upload checks for all three GGUF weight files.

Not established by these checks:

  • Accuracy on structural-engineering benchmarks or actual engineering case assessments.
  • Behavioral equivalence to upstream Ornith or to the MLX releases.
  • GGUF inference throughput, latency, or peak-memory characteristics.
  • Independent functional testing in desktop GUI applications.
  • Engineering-code compliance, calculation reliability, or suitability for safety-critical decisions.

For context, the separate MLX 4-bit release measured 29.46 generated tokens/s in one inference run on an Apple M6 Mac mini. This is not a GGUF benchmark and does not predict performance across GGUF runtimes.

Limitations and responsible use

BuildEng V9 can help organize technical discussions, draft preliminary checklists, and identify missing information. However, it can produce incorrect calculations, omit important load cases, invent code provisions, misinterpret units or material properties, or draw unjustified conclusions from incomplete inputs.

For actual building work, independently verify assumptions, drawings, measurements, loading conditions, applicable codes and standards, and calculations. Decisions affecting structural safety require review by a qualified engineer.

Do not use this model alone to approve structural modifications, diagnose stability, direct temporary works, or authorize site activity.

License and attribution

  • BuildEng V9 fine-tuning and GGUF release: Irfan Uruchi.
  • Base model: Ornith AI โ€” Ornith 1.5 9B, distributed under MIT.
  • Underlying Qwen components: Subject to relevant Apache-2.0 notices.

See the repository's LICENSE and LICENSE-QWEN-APACHE-2.0 files for the relevant license texts. Upstream credit belongs to Ornith AI and the Qwen team. BuildEng V9 is an independent fine-tuned derivative and is not an official upstream release.

Downloads last month
112
GGUF
Model size
9B params
Architecture
qwen35
Hardware compatibility
Log In to add your hardware

4-bit

5-bit

6-bit

Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

Model tree for Irfanuruchi/Ornith-1.5-9B-BuildEng-V9-GGUF

Finetuned
(49)
this model

Dataset used to train Irfanuruchi/Ornith-1.5-9B-BuildEng-V9-GGUF