Qwen3.8-27B-NVFP4-IRIS-Cache-ObjectScript

Fine-tuned Qwen3.8-27B as an InterSystems IRIS / Caché ObjectScript specialist, trained on a task-solving dataset (44% code task→solution, 16% real bugfix pairs, 31% short Q&A, 9% general anti-forgetting) with LR 1e-4 + grad-accum 2 (1200 steps, eval loss 0.966 → 0.787 monotonic), merged to BF16, then re-quantized to NVFP4 (attn FP8 dynamic, MLP NVFP4) with domain-matched ObjectScript calibration.

Won an A/B against the previous v2 fine-tune on real ObjectScript coding prompts (complete working code vs truncated prose).

Model facts

Property	Value
Base model	Qwen/Qwen3.8-27B (hybrid: 48 linear_attn + 16 self_attn layers)
Parameters (BF16)	~27.8B (the HF "18b" label below is the quantized file size in GB, not the parameter count)
Default context window	262,144 tokens (256K) — max_position_embeddings: 262144
Architecture	Qwen3_5ForCausalLM, hidden 5120, 64 layers, 24 heads / 4 KV, vocab 248,320
Quantization	compressed-tensors: NVFP4 (W4A4 static, group 16) on MLP + lm_head, FP8 (W8A8 static) on attention
Quantized size on disk	~18-20 GB (4-bit)

Why it shows "~18B" on Hub: HuggingFace derives the number from the safetensors storage size (~18-20 GB after 4-bit NVFP4 quantization) and displays it as "b". The model is genuinely a 27B-class model (Qwen3.8-27B, ~27.8B BF16 params); 4-bit quantization shrinks the file size, not the parameter count.

Deployment

Serve with vLLM (--quantization compressed-tensors, --max-model-len 262144). Served model name: qwen3.8-27b-iris. Context: 262144 (256K).

On a DGX Spark (GB10/SM121), keep --gpu-memory-utilization <= 0.85 and the freeze guards (--max-cudagraph-capture-size 4, --no-enable-flashinfer-autotune).

Generated with

• llmcompressor 0.13.0 (NVFP4 + FP8 mixed)
• Custom peft + transformers Trainer (BF16 base + 4-bit qLoRA)

Dialect

Trains on InterSystems IRIS / Caché ObjectScript (.cls / .mac / .inc).