Serve Figment v14p through HF ZeroGPU

.env.example

@@ -15,6 +15,12 @@ HF_ENDPOINT_URL=
 OMNI_ENDPOINT_URL=
 HF_TOKEN=
 LLAMA_BASE_URL=http://127.0.0.1:8001/v1
+ZEROGPU_MODEL_REPO=build-small-hackathon/figment-finetuned-model-archive
+ZEROGPU_MODEL_SUBFOLDER=figment_sft_v14p/figment-sft-v14p-lora-merged-bf16
+ZEROGPU_DURATION_SECONDS=120
+ZEROGPU_SIZE=large
+ZEROGPU_MAX_CONTEXT_TOKENS=16384
+ZEROGPU_MAX_GENERATION_TOKENS=1536
 FIGMENT_TRACE_DIR=traces
 FIGMENT_SMOKE_ALLOW_NETWORK=false
 FIGMENT_SMOKE_TIMEOUT_SECONDS=8
@@ -27,6 +33,14 @@ FIGMENT_SMOKE_TRACE_PATH=
 # AUDIO_BACKEND=omni_native
 # ENABLE_AUDIO_INTAKE=true
 
+# Public HF ZeroGPU Space route for the published v14p model:
+# FIGMENT_MODE=hosted
+# MODEL_STACK=local_4b_parakeet
+# MODEL_BACKEND=hf_zerogpu
+# LOCAL_MODEL_ID=figment-sft-v14p-lora-merged-bf16
+# AUDIO_BACKEND=none
+# ENABLE_AUDIO_INTAKE=false
+
 # Local/offline proof. This can use the smaller local model, or self-hosted Omni
 # on adequate local hardware; smoke output records the configured LOCAL_MODEL_ID.
 # FIGMENT_MODE=local

README.md

@@ -7,7 +7,9 @@ sdk: gradio
 sdk_version: 6.17.3
 app_file: app.py
 pinned: false
-python_version: 3.11
+python_version: 3.12.12
+preload_from_hub:
+  - build-small-hackathon/figment-finetuned-model-archive figment_sft_v14p/figment-sft-v14p-lora-merged-bf16/chat_template.jinja,figment_sft_v14p/figment-sft-v14p-lora-merged-bf16/config.json,figment_sft_v14p/figment-sft-v14p-lora-merged-bf16/configuration_nemotron_h.py,figment_sft_v14p/figment-sft-v14p-lora-merged-bf16/generation_config.json,figment_sft_v14p/figment-sft-v14p-lora-merged-bf16/model-00001-of-00002.safetensors,figment_sft_v14p/figment-sft-v14p-lora-merged-bf16/model-00002-of-00002.safetensors,figment_sft_v14p/figment-sft-v14p-lora-merged-bf16/model.safetensors.index.json,figment_sft_v14p/figment-sft-v14p-lora-merged-bf16/modeling_nemotron_h.py,figment_sft_v14p/figment-sft-v14p-lora-merged-bf16/special_tokens_map.json,figment_sft_v14p/figment-sft-v14p-lora-merged-bf16/tokenizer.json,figment_sft_v14p/figment-sft-v14p-lora-merged-bf16/tokenizer_config.json
 ---
 
 # Figment
@@ -22,7 +24,7 @@ Figment turns messy field intake into a card-cited protocol workflow: confirm th
 
 | Surface | Current evidence | What it means | Boundary |
 | --- | --- | --- | --- |
-| Public Space | [build-small-hackathon/figment](https://huggingface.co/spaces/build-small-hackathon/figment) is `RUNNING` and served HTTP 200 when checked on 2026-06-14 at Space commit `ea4d971be2b8d4e1841648e5a9adc5142d16a299`. | The Gradio app is deployable and reachable. | The public no-secret route may use the labeled canned fallback; that is deployment proof, not live hosted-model proof. |
+| Public Space | [build-small-hackathon/figment](https://huggingface.co/spaces/build-small-hackathon/figment) is configured for `MODEL_BACKEND=hf_zerogpu` against the published v14p BF16 artifact. | The public route can call the same 4B v14p model archive used in the strongest small-model evidence bundle. | ZeroGPU has queue/quota limits; live route proof should include a fresh Space call and trace route. |
 | Hosted Omni eval | `31/50` whole-output competence, `8/50` full fallback, `480/650` model-retained fields, `170/650` deterministic patches, and `50/50` final validation. | Hosted Omni can carry bounded fields, and the app can keep outputs inside the safety contract. | `50/50` final validation is app safety after validation, repair, and fallback. It is not pure model performance. |
 | 4B LoRA system eval | v14p repair-union on the corrected 150-case field-workflow holdout: `150/150` competence, `150/150` expected labels, `150/150` final validation, `0` deterministic patches, `0` fallback. Raw first-pass success is `146/150`; `4/150` cases close through focused model repair. | The strongest documented small-model result is model-owned output plus model repair on a synthetic/de-identified holdout. | This is not clinical validation, target-user validation, local ASR proof, or proof that raw first-pass output solved every case. |
 | Public artifacts | [model archive](https://huggingface.co/build-small-hackathon/figment-finetuned-model-archive) and [eval/training dataset](https://huggingface.co/datasets/build-small-hackathon/figment-eval-traces). | Versioned BF16/GGUF model artifacts, synthetic corpora, eval traces, and summaries are inspectable outside this checkout. | Generated `traces/`, `data/finetune/`, weights, and checkpoint folders are intentionally not part of a clean clone. |
@@ -69,7 +71,7 @@ app.py
   -> figment/rules.py              deterministic danger-sign rules
   -> figment/retrieval.py          local protocol-card retrieval
   -> figment/prompt_builder.py     bounded navigator prompt
-  -> figment/model_client.py       hosted Omni, local OpenAI-compatible, or canned route
+  -> figment/model_client.py       hosted Omni, HF ZeroGPU, local OpenAI-compatible, or canned route
   -> figment/navigator.py          raw output, scaffold, repair, fallback orchestration
   -> figment/validators.py         schema, citations, urgency floor, safety checks
   -> figment/field_provenance.py   model_raw / model_repaired / deterministic_fallback labels
@@ -87,12 +89,13 @@ The safety pattern is deliberate:
 
 ## Models
 
-Figment supports three runtime routes:
+Figment supports four runtime routes:
 
 | Route | Backend | Use |
 | --- | --- | --- |
 | Canned fallback | `MODEL_BACKEND=canned` | No-secret app smoke, UI development, honest fallback traces. |
 | Hosted Omni | `MODEL_BACKEND=hosted_omni` with `NVIDIA_API_KEY` | Live hosted demo and hosted eval path using `nvidia/nemotron-3-nano-omni-30b-a3b-reasoning`. |
+| HF ZeroGPU v14p | `MODEL_BACKEND=hf_zerogpu` with `ZEROGPU_MODEL_REPO` / `ZEROGPU_MODEL_SUBFOLDER` | Public Space route using the published v14p BF16 merged model on Hugging Face ZeroGPU. |
 | Local OpenAI-compatible | `MODEL_BACKEND=llama_cpp` with `LLAMA_BASE_URL` | Local text-navigation route for the 4B BF16/GGUF artifacts and local evidence bundles. |
 
 The Build Small constraint is <=32B total parameters. The hosted Omni path is tracked with a parameter-count caveat: the NVIDIA model-card body reports 31B total parameters, while sidebar counts have differed. The 4B BF16 base model, `nvidia/NVIDIA-Nemotron-3-Nano-4B-BF16`, is the local text-navigation training target.

app.py

@@ -344,7 +344,7 @@ def _app_header_html() -> str:
 
 def _statusline_html(config: FigmentConfig) -> str:
     audio_chip = "green" if config.enable_audio_intake else "amber"
-    backend_chip = "blue" if config.model_backend == "hosted_omni" else "amber"
+    backend_chip = "blue" if config.model_backend in {"hosted_omni", "hf_zerogpu"} else "amber"
     return f"""
     <div class="figment-statusline">
       <strong>Runtime</strong>
@@ -955,6 +955,8 @@ __FIGMENT_CSS__
 def _model_mode_label(config: FigmentConfig) -> str:
     if config.model_backend == "hosted_omni":
         return "Configured backend: hosted_omni"
+    if config.model_backend == "hf_zerogpu":
+        return "Configured backend: hf_zerogpu"
     if config.model_backend == "llama_cpp":
         return "Configured backend: llama_cpp"
     return "Configured backend: canned"

figment/config.py

@@ -15,9 +15,12 @@ PARAKEET_ASR_MODEL_ID = "nvidia/parakeet-rnnt-1.1b"
 NVIDIA_OMNI_API_MODEL_ID = "nvidia/nemotron-3-nano-omni-30b-a3b-reasoning"
 NVIDIA_API_BASE_URL = "https://integrate.api.nvidia.com/v1"
 FIGMENT_CANNED_MODEL_ID = "figment-canned-deterministic"
+FIGMENT_V14P_MODEL_REPO = "build-small-hackathon/figment-finetuned-model-archive"
+FIGMENT_V14P_MODEL_SUBFOLDER = "figment_sft_v14p/figment-sft-v14p-lora-merged-bf16"
+FIGMENT_V14P_MODEL_ID = "figment-sft-v14p-lora-merged-bf16"
 
 MODEL_STACKS = {"omni_native", "local_4b_parakeet"}
-MODEL_BACKENDS = {"hosted_omni", "llama_cpp", "canned"}
+MODEL_BACKENDS = {"hosted_omni", "llama_cpp", "hf_zerogpu", "canned"}
 AUDIO_BACKENDS = {"omni_native", "parakeet_nemo", "canned", "none"}
 FIGMENT_MODES = {"hosted", "local", "canned"}
 
@@ -86,6 +89,8 @@ class FigmentConfig:
     omni_endpoint_url: str = ""
     hf_endpoint_url: str = ""
     llama_base_url: str = "http://127.0.0.1:8001/v1"
+    zerogpu_model_repo: str = FIGMENT_V14P_MODEL_REPO
+    zerogpu_model_subfolder: str = FIGMENT_V14P_MODEL_SUBFOLDER
     trace_dir: Path = Path("traces")
 
     @classmethod
@@ -109,7 +114,7 @@ class FigmentConfig:
             backend = {"hosted": "hosted_omni", "local": "llama_cpp", "canned": "canned"}[backend]
         stack = os.getenv("MODEL_STACK", "").strip()
         if not stack:
-            stack = "local_4b_parakeet" if backend == "llama_cpp" else "omni_native"
+            stack = "local_4b_parakeet" if backend in {"llama_cpp", "hf_zerogpu"} else "omni_native"
         audio = os.getenv("AUDIO_BACKEND", "none").strip() or "none"
         allow_self_hosted_omni = _bool_env(os.getenv("ALLOW_SELF_HOSTED_OMNI"), False)
         mode_errors = _mode_consistency_errors(
@@ -139,11 +144,18 @@ class FigmentConfig:
             nvidia_api_key=nvidia_api_key,
             nvidia_base_url=os.getenv("NVIDIA_BASE_URL", NVIDIA_API_BASE_URL).strip() or NVIDIA_API_BASE_URL,
             nvidia_model_id=os.getenv("NVIDIA_MODEL_ID", NVIDIA_OMNI_API_MODEL_ID).strip() or NVIDIA_OMNI_API_MODEL_ID,
-            local_model_id=os.getenv("LOCAL_MODEL_ID", NVIDIA_NEMOTRON_3_NANO_4B_BF16_MODEL_ID).strip()
-            or NVIDIA_NEMOTRON_3_NANO_4B_BF16_MODEL_ID,
+            local_model_id=os.getenv(
+                "LOCAL_MODEL_ID",
+                FIGMENT_V14P_MODEL_ID if backend == "hf_zerogpu" else NVIDIA_NEMOTRON_3_NANO_4B_BF16_MODEL_ID,
+            ).strip()
+            or (FIGMENT_V14P_MODEL_ID if backend == "hf_zerogpu" else NVIDIA_NEMOTRON_3_NANO_4B_BF16_MODEL_ID),
             omni_endpoint_url=omni_endpoint_url,
             hf_endpoint_url=hf_endpoint_url,
             llama_base_url=os.getenv("LLAMA_BASE_URL", "http://127.0.0.1:8001/v1").strip(),
+            zerogpu_model_repo=os.getenv("ZEROGPU_MODEL_REPO", FIGMENT_V14P_MODEL_REPO).strip()
+            or FIGMENT_V14P_MODEL_REPO,
+            zerogpu_model_subfolder=os.getenv("ZEROGPU_MODEL_SUBFOLDER", FIGMENT_V14P_MODEL_SUBFOLDER).strip()
+            or FIGMENT_V14P_MODEL_SUBFOLDER,
             trace_dir=Path(os.getenv("FIGMENT_TRACE_DIR", "traces").strip() or "traces"),
         ).validated()
 
@@ -159,6 +171,8 @@ class FigmentConfig:
             errors.append("MODEL_BACKEND=hosted_text_nemotron is retired; use MODEL_BACKEND=llama_cpp for local_4b_parakeet")
         elif self.model_backend not in MODEL_BACKENDS:
             errors.append(f"MODEL_BACKEND must be one of {sorted(MODEL_BACKENDS)}")
+        if self.model_backend == "hf_zerogpu" and self.model_stack != "local_4b_parakeet":
+            errors.append("MODEL_BACKEND=hf_zerogpu requires MODEL_STACK=local_4b_parakeet")
         if self.audio_backend not in AUDIO_BACKENDS:
             errors.append(f"AUDIO_BACKEND must be one of {sorted(AUDIO_BACKENDS)}")
         if self.audio_backend == "parakeet_nemo" and not self.allow_local_asr:
@@ -188,7 +202,7 @@ class FigmentConfig:
     def active_model_id(self) -> str:
         if self.model_backend == "hosted_omni":
             return self.nvidia_model_id
-        if self.model_backend == "llama_cpp" or self.model_stack == "local_4b_parakeet":
+        if self.model_backend in {"llama_cpp", "hf_zerogpu"} or self.model_stack == "local_4b_parakeet":
             return self.local_model_id
         return FIGMENT_CANNED_MODEL_ID
 

figment/model_client.py

@@ -15,6 +15,7 @@ import wave
 from typing import Any
 
 from .config import FigmentConfig, load_config
+from .zerogpu_runtime import generate_zero_gpu_json
 
 
 class ModelClientError(RuntimeError):
@@ -126,6 +127,13 @@ class ModelClient:
                 model_id=self.config.local_model_id,
                 auth_headers={},
             )
+        if self.config.model_backend == "hf_zerogpu":
+            return generate_zero_gpu_json(
+                prompt=prompt,
+                model_repo=self.config.zerogpu_model_repo,
+                model_subfolder=self.config.zerogpu_model_subfolder,
+                model_id=self.config.local_model_id,
+            )
         if self.config.model_backend == "hosted_omni":
             endpoint = self.config.omni_endpoint_url or self.config.hf_endpoint_url or self.config.nvidia_base_url
             if not endpoint:

figment/trace.py

@@ -135,7 +135,7 @@ def _derive_final_route(
         return "model_repaired"
     if fallback_tier == "canned":
         return "canned_backend"
-    if raw_route in {"hosted_omni", "llama_cpp"}:
+    if raw_route in {"hosted_omni", "llama_cpp", "hf_zerogpu"}:
         return "live_model_generated"
     return raw_route or "unknown"
 

figment/zerogpu_runtime.py

@@ -0,0 +1,147 @@
+"""ZeroGPU Transformers runtime for Figment's published v14p model."""
+
+from __future__ import annotations
+
+import json
+import os
+import threading
+from typing import Any
+
+
+try:  # pragma: no cover - exercised in Hugging Face Spaces.
+    import spaces
+except ImportError:  # pragma: no cover - local tests use the no-op decorator.
+
+    class _SpacesCompat:
+        @staticmethod
+        def GPU(*_: Any, **__: Any) -> Any:  # noqa: N802 - mirrors the spaces API.
+            def decorator(function: Any) -> Any:
+                return function
+
+            return decorator
+
+    spaces = _SpacesCompat()
+
+
+_RUNTIMES: dict[tuple[str, str], "_ZeroGpuRuntime"] = {}
+_RUNTIME_LOCK = threading.Lock()
+
+
+@spaces.GPU(
+    duration=int(os.getenv("ZEROGPU_DURATION_SECONDS", "120")),
+    size=os.getenv("ZEROGPU_SIZE", "large"),
+)
+def generate_zero_gpu_json(
+    *,
+    prompt: str,
+    model_repo: str,
+    model_subfolder: str,
+    model_id: str,
+) -> dict[str, Any]:
+    runtime = _runtime_for(model_repo, model_subfolder, model_id)
+    return runtime.generate_json(prompt)
+
+
+def _runtime_for(model_repo: str, model_subfolder: str, model_id: str) -> "_ZeroGpuRuntime":
+    key = (model_repo, model_subfolder)
+    with _RUNTIME_LOCK:
+        runtime = _RUNTIMES.get(key)
+        if runtime is None:
+            runtime = _ZeroGpuRuntime(model_repo=model_repo, model_subfolder=model_subfolder, model_id=model_id)
+            _RUNTIMES[key] = runtime
+        return runtime
+
+
+class _ZeroGpuRuntime:
+    def __init__(self, *, model_repo: str, model_subfolder: str, model_id: str) -> None:
+        import torch
+        from transformers import AutoModelForCausalLM
+        from transformers import AutoTokenizer
+
+        self.model_id = model_id
+        self.max_context_tokens = int(os.getenv("ZEROGPU_MAX_CONTEXT_TOKENS", "16384"))
+        self.max_generation_tokens = int(os.getenv("ZEROGPU_MAX_GENERATION_TOKENS", "1536"))
+        self.lock = threading.Lock()
+        self.torch = torch
+        self.tokenizer = AutoTokenizer.from_pretrained(
+            model_repo,
+            subfolder=model_subfolder,
+            trust_remote_code=True,
+        )
+        if self.tokenizer.pad_token is None:
+            self.tokenizer.pad_token = self.tokenizer.eos_token
+        self.model = AutoModelForCausalLM.from_pretrained(
+            model_repo,
+            subfolder=model_subfolder,
+            trust_remote_code=True,
+            torch_dtype=torch.bfloat16,
+        )
+        self.model.to("cuda")
+        self.model.eval()
+        self.model.config.use_cache = True
+
+    def generate_json(self, prompt: str) -> dict[str, Any]:
+        messages = [{"role": "user", "content": prompt}]
+        with self.lock:
+            input_ids = self.tokenizer.apply_chat_template(
+                messages,
+                add_generation_prompt=True,
+                return_tensors="pt",
+            )
+            device = next(self.model.parameters()).device
+            input_ids = input_ids.to(device)
+            input_len = int(input_ids.shape[-1])
+            available_tokens = max(1, self.max_context_tokens - input_len - 8)
+            max_new_tokens = max(1, min(self.max_generation_tokens, available_tokens))
+            attention_mask = self.torch.ones_like(input_ids)
+            with self.torch.inference_mode():
+                output_ids = self.model.generate(
+                    input_ids=input_ids,
+                    attention_mask=attention_mask,
+                    max_new_tokens=max_new_tokens,
+                    do_sample=False,
+                    pad_token_id=self.tokenizer.pad_token_id,
+                    eos_token_id=self.tokenizer.eos_token_id,
+                )
+            generated_ids = output_ids[0, input_len:]
+            content = self.tokenizer.decode(generated_ids, skip_special_tokens=True).strip()
+        return _parse_json_object(content)
+
+
+def _parse_json_object(content: str) -> dict[str, Any]:
+    text = content.strip()
+    try:
+        parsed = json.loads(text)
+    except json.JSONDecodeError:
+        decoder = json.JSONDecoder()
+        for index, char in enumerate(text):
+            if char != "{":
+                continue
+            try:
+                parsed, _ = decoder.raw_decode(text[index:])
+                break
+            except json.JSONDecodeError:
+                continue
+        else:
+            raise
+    if isinstance(parsed, dict):
+        return parsed
+    raise json.JSONDecodeError("model response JSON was not an object", text, 0)
+
+
+def _preload_default_runtime_from_env() -> None:
+    if os.getenv("MODEL_BACKEND") != "hf_zerogpu":
+        return
+    if os.getenv("FIGMENT_SKIP_ZEROGPU_PRELOAD", "").strip().lower() in {"1", "true", "yes", "on"}:
+        return
+    model_repo = os.getenv("ZEROGPU_MODEL_REPO", "build-small-hackathon/figment-finetuned-model-archive").strip()
+    model_subfolder = os.getenv(
+        "ZEROGPU_MODEL_SUBFOLDER",
+        "figment_sft_v14p/figment-sft-v14p-lora-merged-bf16",
+    ).strip()
+    model_id = os.getenv("LOCAL_MODEL_ID", "figment-sft-v14p-lora-merged-bf16").strip()
+    if model_repo and model_subfolder and model_id:
+        _runtime_for(model_repo, model_subfolder, model_id)
+
+
+_preload_default_runtime_from_env()

requirements.txt

@@ -5,3 +5,15 @@ httpx>=0.27,<1
 pydantic>=2.7,<3
 python-dotenv>=1.0,<2
 PyYAML>=6.0,<7
+spaces>=0.40,<1
+torch>=2.8,<2.12
+transformers>=4.52,<5
+accelerate>=1.8,<2
+einops>=0.8,<1
+protobuf>=5,<7
+safetensors>=0.5,<1
+sentencepiece>=0.2,<1
+setuptools>=70
+wheel>=0.45
+causal-conv1d>=1.6,<2
+mamba-ssm>=2.2,<3

tests/test_model_client.py

@@ -145,6 +145,38 @@ def test_local_llama_cpp_uses_local_base_url_without_hosted_auth(monkeypatch: An
     assert captured["body"]["model"] == "local-test-model"
 
 
+def test_hf_zerogpu_backend_uses_published_v14p_runtime(monkeypatch: Any) -> None:
+    captured: dict[str, Any] = {}
+
+    def fake_generate_zero_gpu_json(*, prompt: str, model_repo: str, model_subfolder: str, model_id: str) -> dict[str, Any]:
+        captured["prompt"] = prompt
+        captured["model_repo"] = model_repo
+        captured["model_subfolder"] = model_subfolder
+        captured["model_id"] = model_id
+        return {"protocol_urgency": "monitor", "source_cards": []}
+
+    monkeypatch.setattr("figment.model_client.generate_zero_gpu_json", fake_generate_zero_gpu_json)
+
+    client = ModelClient(
+        FigmentConfig(
+            model_backend="hf_zerogpu",
+            model_stack="local_4b_parakeet",
+            local_model_id="figment-sft-v14p-lora-merged-bf16",
+            zerogpu_model_repo="build-small-hackathon/figment-finetuned-model-archive",
+            zerogpu_model_subfolder="figment_sft_v14p/figment-sft-v14p-lora-merged-bf16",
+        )
+    )
+    result = client.generate_json("Return JSON.", {})
+
+    assert result["protocol_urgency"] == "monitor"
+    assert captured == {
+        "prompt": "Return JSON.",
+        "model_repo": "build-small-hackathon/figment-finetuned-model-archive",
+        "model_subfolder": "figment_sft_v14p/figment-sft-v14p-lora-merged-bf16",
+        "model_id": "figment-sft-v14p-lora-merged-bf16",
+    }
+
+
 def test_custom_hf_endpoint_prefers_hf_token_over_nvidia_key(monkeypatch: Any) -> None:
     captured: dict[str, Any] = {}
 

tests/test_runtime_honesty.py

@@ -12,10 +12,12 @@ CONFIG_ENV_KEYS = (
     "MODEL_STACK",
     "MODEL_BACKEND",
     "AUDIO_BACKEND",
-        "ENABLE_AUDIO_INTAKE",
-        "ALLOW_LOCAL_ASR",
-        "ALLOW_SELF_HOSTED_OMNI",
-        "ALLOW_STRETCH_STACK",
+    "ENABLE_AUDIO_INTAKE",
+    "ALLOW_LOCAL_ASR",
+    "ALLOW_SELF_HOSTED_OMNI",
+    "ALLOW_STRETCH_STACK",
+    "ZEROGPU_MODEL_REPO",
+    "ZEROGPU_MODEL_SUBFOLDER",
     "HF_MODEL_ID",
     "NVIDIA_MODEL_ID",
     "NVIDIA_BASE_URL",
@@ -73,10 +75,27 @@ def test_env_config_allows_explicit_local_self_hosted_omni(monkeypatch: pytest.M
     assert config.active_model_id == "nvidia/Nemotron-3-Nano-Omni-30B-A3B-Reasoning-BF16"
 
 
+def test_env_config_zerogpu_backend_defaults_to_v14p(monkeypatch: pytest.MonkeyPatch, tmp_path) -> None:
+    monkeypatch.chdir(tmp_path)
+    _clear_config_env(monkeypatch)
+    monkeypatch.setenv("FIGMENT_MODE", "hosted")
+    monkeypatch.setenv("MODEL_BACKEND", "hf_zerogpu")
+
+    config = FigmentConfig.from_env()
+
+    assert config.model_backend == "hf_zerogpu"
+    assert config.model_stack == "local_4b_parakeet"
+    assert config.local_model_id == "figment-sft-v14p-lora-merged-bf16"
+    assert config.zerogpu_model_repo == "build-small-hackathon/figment-finetuned-model-archive"
+    assert config.zerogpu_model_subfolder == "figment_sft_v14p/figment-sft-v14p-lora-merged-bf16"
+    assert config.active_model_id == "figment-sft-v14p-lora-merged-bf16"
+
+
 @pytest.mark.parametrize(
     ("model_route", "events", "expected_final_route"),
     [
         ({"model_backend": "hosted_omni", "fallback_reason": None}, [], "live_model_generated"),
+        ({"model_backend": "hf_zerogpu", "fallback_reason": None}, [], "live_model_generated"),
         ({"model_backend": "hosted_omni", "fallback_reason": None}, ["navigator output repaired by hosted retry"], "model_repaired"),
         (
             {"model_backend": "hosted_omni", "fallback_reason": None, "field_level_fallback_used": True},