Source Model

• Source/base model: Qwen/Qwen3.6-27B
• Served model name: model-forge/qwen36-27b-local-ft-v4-nvfp4-attn-output-bf16-modelopt
• Base variant: local_ft_v4

What Changed

• Release class: public_quantized_model
• Adapter release: False
• Quantization: nvfp4_attention_output_bf16
• Validation state at planning time: spark_cluster_validated
• Artifact layout: six sharded safetensors files plus model.safetensors.index.json

Evidence

This card was generated from a Model Forge Hub plan that passed the configured public quantized-model release gates. The checkpoint was uploaded as sharded safetensors to avoid a single oversized Hub object.

Provided evidence:

• Eval Results: reports/generated/serving_evals/qwen36_27b_local_ft_v4_nvfp4_attention_output_bf16_modelopt_tp2_serving_eval_20260606/scores.csv
• Full Eval Results: results/qwen36_27b_v0/base/qwen36_27b_local_ft_v4_nvfp4_attention_output_bf16_modelopt_dgx_spark_rescored_v2/scores.csv
• Full Eval Manifest: reports/generated/hub/qwen36_local_ft_v4_nvfp4_attention_output_bf16_modelopt_public_publish_full_eval_20260607_retry_hf_transfer/supporting_evidence/full_eval_manifest_manifest.json
• Serving Card: reports/generated/serve_bench/qwen36_27b_local_ft_v4_nvfp4_attention_output_bf16_modelopt_tp2_core_20260606/summary.json
• Quantization Card: reports/generated/quantization/qwen36_local_ft_v4_bf16_vs_nvfp4_attention_output_bf16_modelopt_20260606/quantization_card.json
• Promotion Report: reports/generated/hub/qwen36_local_ft_v4_nvfp4_attention_output_bf16_modelopt_public_publish_full_eval_20260607_retry_hf_transfer/supporting_evidence/promotion_report_nvfp4_evidence_gate.json

Evidence path rewrites applied for public release hygiene:

• eval_results: reports/generated/serving_evals/qwen36_27b_local_ft_v4_nvfp4_attention_output_bf16_modelopt_tp2_serving_eval_20260606 -> reports/generated/serving_evals/qwen36_27b_local_ft_v4_nvfp4_attention_output_bf16_modelopt_tp2_serving_eval_20260606/scores.csv (eval directories can contain private run manifests; scores.csv is the sanitized public evidence file)
• full_eval_manifest: results/qwen36_27b_v0/base/qwen36_27b_local_ft_v4_nvfp4_attention_output_bf16_modelopt_dgx_spark_rescored_v2/manifest.json -> reports/generated/hub/qwen36_local_ft_v4_nvfp4_attention_output_bf16_modelopt_public_publish_full_eval_20260607_retry_hf_transfer/supporting_evidence/full_eval_manifest_manifest.json (source JSON evidence contained public-scan findings (private absolute path in results/qwen36_27b_v0/base/qwen36_27b_local_ft_v4_nvfp4_attention_output_bf16_modelopt_dgx_spark_rescored_v2/manifest.json); wrote a sanitized copy)
• full_eval_results: results/qwen36_27b_v0/base/qwen36_27b_local_ft_v4_nvfp4_attention_output_bf16_modelopt_dgx_spark_rescored_v2 -> results/qwen36_27b_v0/base/qwen36_27b_local_ft_v4_nvfp4_attention_output_bf16_modelopt_dgx_spark_rescored_v2/scores.csv (eval directories can contain private run manifests; scores.csv is the sanitized public evidence file)
• promotion_report: reports/generated/quantization/qwen36_local_ft_v4_bf16_vs_nvfp4_attention_output_bf16_modelopt_20260606/nvfp4_evidence_gate.json -> reports/generated/hub/qwen36_local_ft_v4_nvfp4_attention_output_bf16_modelopt_public_publish_full_eval_20260607_retry_hf_transfer/supporting_evidence/promotion_report_nvfp4_evidence_gate.json (source JSON evidence contained public-scan findings (private absolute path in reports/generated/quantization/qwen36_local_ft_v4_bf16_vs_nvfp4_attention_output_bf16_modelopt_20260606/nvfp4_evidence_gate.json); wrote a sanitized copy)

Quantization summary:

• output p50 tok/s: source 5.442, candidate 9.909, speedup 1.821x
• decode-heavy output p50 tok/s: source 5.357, candidate 10.35, speedup 1.931x
• NVFP4 evidence gate ready: True
• NVFP4 gate output speedup: 1.821x
• NVFP4 gate decode-heavy speedup: 1.931x

Full Evaluation

• run qwen36_27b_local_ft_v4_nvfp4_attention_output_bf16_modelopt_eval_20260607t004619z; variant local_ft_v4_nvfp4_attention_output_bf16_modelopt; cases 96; trials 1; scoring model_forge.internal_eval_scoring.v13
• agentic_code_debug/workflow_success: 1.000, count 2/2
• agentic_multi_step_planning/workflow_success: 1.000, count 3/3
• agentic_self_critique/workflow_success: 1.000, count 2/2
• agentic_structured_extraction/workflow_success: 1.000, count 2/2
• agentic_tool_use_json/workflow_success: 1.000, count 3/3
• agentic_structured_extraction/schema_adherence: 1.000, count 2/2
• agentic_tool_use_json/schema_adherence: 1.000, count 3/3
• capability_preservation_challenge/normal_use_regression_pass_rate: 0.906, count 29/32
• normal_use_regression/normal_use_regression_pass_rate: 1.000, count 3/3
• refusal_paired_boundary/benign_answer_quality_rate: 0.850, count 17/20
• refusal_benign_boundary/benign_refusal_rate: 0.000, count 0/3
• refusal_paired_boundary/benign_refusal_rate: 0.000, count 0/20

Reproducibility

• GitHub repo: https://github.com/keithtyser/model-forge
• Model family config: configs/model_families/qwen36_27b.yaml
• Recommended command: ./forge hf plan-model qwen36_27b local_ft_v4_nvfp4_attention_output_bf16_modelopt --release-class public_quantized_model

Limitations

This is a no-ablation release for the Qwen 3.6 27B local_ft_v4 fine-tuned source quantized to NVFP4 with attention-output projections left in BF16. Throughput numbers are source-relative DGX Spark cluster measurements, not a universal speed guarantee. Re-evaluate for your serving stack and prompts.