gemma-4-e2b-flowcast-v4-cascade-qwen · flowcast-v4-cascade-qwen

Flowcast v4 Track B2 is a cascade architecture for size-constrained deploys: a Qwen3-4B light adapter handles intent classification and dictation polish; flowcast-sota-v3 lazy-loads for automation planning.

Small hot path. Full automation when you need it.

Benchmarks (vs v3)

Gate	v3	B2 cascade	Δ
Core hard quality (117)	100%	100%	—
Expanded hard quality (170)	99.4%	99.4%	—
Core p50 latency	~1002ms	~999ms	~same
Hot path disk	2.5 GB	~2.26 GB	Qwen only

Distilled from v3 teacher labels on intent cases (Slack writable-focus fix).

Quick start (cascade)

pip install mlx-lm huggingface_hub

from huggingface_hub import snapshot_download
from gemmaflow_tune.production import create_cascade_runner

light = snapshot_download("nsalerni/gemma-4-e2b-flowcast-v4-cascade-qwen")
heavy = snapshot_download("nsalerni/gemma-4-e2b-flowcast-v3")

runner = create_cascade_runner(
    light_adapter_path=light,
    heavy_adapter_path=heavy,
)
runner.load()

# Intent/dictation routes to Qwen; automation lazy-loads v3 E2B
result = runner.generate("", user_content="...", suite="intent_classification", max_tokens=64)
print(result.text)

Files

File	Description
adapters.safetensors	Qwen3 light LoRA weights
adapter_config.json	LoRA config + heavy adapter reference
inference_config.json	Cascade runtime settings + benchmark scores

Recommended inference settings

{
  "runner_kind": "cascade",
  "prompt_mode": "hybrid_slim",
  "json_early_stop": true,
  "use_prompt_kv_cache": true,
  "temperature": 0.02,
  "top_p": 0.85
}

License

Apache 2.0. Base models subject to their respective licenses.