diff --git a/studio/backend/core/inference/diffusion.py b/studio/backend/core/inference/diffusion.py
index 3c22b25d32..364df883ec 100644
--- a/studio/backend/core/inference/diffusion.py
+++ b/studio/backend/core/inference/diffusion.py
@@ -48,6 +48,7 @@ from .diffusion_device import (
     resolve_diffusion_device_target,
 )
 from .diffusion_ideogram4 import ideogram4_repo_is_fp8, load_ideogram4_pipeline
+from .diffusion_hidream import HIDREAM_FAMILY_NAME, hidream_te4_kwargs
 from .diffusion_krea2 import KREA2_FAMILY_NAME, load_krea2_pipeline
 from .diffusion_memory import (
     MEMORY_MODE_BALANCED,
@@ -302,6 +303,12 @@ _TRUSTED_NON_GGUF_REPOS = frozenset(
         # HunyuanImage 2.1: the community diffusers mirror (open, tencent-hunyuan-community
         # license), safetensors-only, including the diffusers-native guider components.
         "hunyuanvideo-community/hunyuanimage-2.1-diffusers",
+        # HiDream-I1: open MIT-weights repos, all three variants one family. The Llama TE the
+        # model_index names comes from the unsloth mirror (diffusion_hidream.py), which the
+        # unsloth/ org prefix already trusts.
+        "hidream-ai/hidream-i1-full",
+        "hidream-ai/hidream-i1-dev",
+        "hidream-ai/hidream-i1-fast",
         # Ideogram 4: no bf16 ships. -fp8 stores the two DiTs as raw float8 (the family base);
         # the two nf4 repos are identical bnb-4bit exports (both listed so either id loads).
         "ideogram-ai/ideogram-4-fp8",
@@ -1388,6 +1395,10 @@ class DiffusionBackend:
                             pipe_kwargs: dict[str, Any] = {"torch_dtype": dtype}
                             if hf_token:
                                 pipe_kwargs["token"] = hf_token
+                            if fam.name == HIDREAM_FAMILY_NAME:
+                                # The repo names a Llama text_encoder_4 it does not ship;
+                                # supply it from the open mirror (diffusion_hidream.py).
+                                pipe_kwargs.update(hidream_te4_kwargs(dtype, hf_token))
                             # The prefetched snapshot dir keeps from_pretrained off the hub (its
                             # sweep re-pulls files the scoped prefetch skipped: 24 GB per FLUX.1).
                             pipe = pipeline_cls.from_pretrained(
@@ -1428,6 +1439,9 @@ class DiffusionBackend:
                             pipe_kwargs = {"torch_dtype": dtype, "transformer": transformer}
                             if hf_token:
                                 pipe_kwargs["token"] = hf_token
+                            if fam.name == HIDREAM_FAMILY_NAME:
+                                # Same Llama TE4 assembly as the full-pipeline branch above.
+                                pipe_kwargs.update(hidream_te4_kwargs(dtype, hf_token))
                             pipe = pipeline_cls.from_pretrained(
                                 _base_local_dir or base, **pipe_kwargs
                             )
@@ -1864,6 +1878,10 @@ class DiffusionBackend:
         pipe_kwargs: dict[str, Any] = {"torch_dtype": dtype, "transformer": transformer}
         if hf_token:
             pipe_kwargs["token"] = hf_token
+        if getattr(fam, "name", None) == HIDREAM_FAMILY_NAME:
+            # The repo ships no Llama text_encoder_4; assemble it from the open mirror
+            # (diffusion_hidream.py) exactly like the full-pipeline load branch.
+            pipe_kwargs.update(hidream_te4_kwargs(dtype, hf_token))
         pipe = pipeline_cls.from_pretrained(base_local_dir or base, **pipe_kwargs)
         pipe.to(device)
         return pipe
diff --git a/studio/backend/core/inference/diffusion_auto_policy.py b/studio/backend/core/inference/diffusion_auto_policy.py
index 851ea65115..32e3a3ee02 100644
--- a/studio/backend/core/inference/diffusion_auto_policy.py
+++ b/studio/backend/core/inference/diffusion_auto_policy.py
@@ -54,6 +54,10 @@ _FAMILY_BF16_GB: dict[str, tuple[float, float, float]] = {
     "lumina-2": (5.2, 5.2, 0.2),
     # 17B dual-stream DiT (32.5 GB bf16 on disk) + Qwen2.5-VL 15.5 GB + ByT5 0.8 GB.
     "hunyuanimage-2.1": (32.5, 16.3, 0.8),
+    # 17B MoE DiT (34.2 GB bf16) + FOUR text encoders: CLIP-L 0.5 + CLIP-G 2.8 + T5-XXL 9.5
+    # from the repo, plus the Llama-3.1-8B text_encoder_4 (~16 GB bf16) assembled from the
+    # open mirror at load time (diffusion_hidream.py).
+    "hidream-i1": (34.2, 28.8, 0.2),
     # Two ~9.3B DiTs (conditional + unconditional_transformer for Ideogram's dual-branch CFG),
     # both resident, plus a Qwen3-VL encoder. The vendor stores them as raw float8; these are the
     # bf16-resident sizes after the dtype cast, so each doubles (37.2 = 2 x 18.6, encoder 16.3).
diff --git a/studio/backend/core/inference/diffusion_families.py b/studio/backend/core/inference/diffusion_families.py
index 87b897d88f..6cdb6c6685 100644
--- a/studio/backend/core/inference/diffusion_families.py
+++ b/studio/backend/core/inference/diffusion_families.py
@@ -347,6 +347,21 @@ _FAMILIES: tuple[DiffusionFamily, ...] = (
         # Exported bf16-only; keep the fp16 fallback off like z-image / krea-2.
         fp16_incompatible = True,
     ),
+    # HiDream-I1: a 17B MoE DiT (16 double + 32 single layers, 4 routed experts) with FOUR text
+    # encoders. The repos ship CLIP-L/CLIP-G/T5-XXL but NOT the Llama-3.1-8B text_encoder_4 their
+    # model_index names: the loader assembles it from the open unsloth mirror
+    # (diffusion_hidream.py). Full / Dev / Fast share the arch, so one family covers all three
+    # (per-variant step/guidance defaults below). city96 publishes a GGUF but the GGUF path would
+    # need the same TE4 assembly for tiny demand, so no GGUF artifact is wired yet.
+    DiffusionFamily(
+        name = "hidream-i1",
+        pipeline_class = "HiDreamImagePipeline",
+        transformer_class = "HiDreamImageTransformer2DModel",
+        base_repo = "HiDream-ai/HiDream-I1-Full",
+        aliases = ("hidream", "hidream-i1-full", "hidream-i1-dev", "hidream-i1-fast"),
+        # Exported bf16-only; keep the fp16 fallback off like the other modern DiTs.
+        fp16_incompatible = True,
+    ),
     # Ideogram 4 (diffusers >= 0.39): a 34-layer DiT PAIR (conditional + unconditional_transformer
     # for dual-branch CFG, both ~9B, so memory planning counts two DiTs) with a Qwen3-VL encoder.
     # No bf16 checkpoint: ideogram-4-fp8 (raw float8, upcast on load) is the highest-precision
@@ -532,6 +547,11 @@ _GENERATION_DEFAULTS: tuple[tuple[str, int, float], ...] = (
     # HunyuanImage 2.1 model-card: 50 steps; the guidance value feeds the call's
     # distilled_guidance_scale (default 3.25), while real CFG runs inside the repo guiders.
     ("hunyuanimage", 50, 3.25),
+    # HiDream-I1 upstream inference.py: Full 50 steps / guidance 5; the distilled Dev (28) and
+    # Fast (16) run guidance-free. Specific keys precede the generic "hidream" (Full + fallback).
+    ("hidream-i1-dev", 28, 0.0),
+    ("hidream-i1-fast", 16, 0.0),
+    ("hidream", 50, 5.0),
     # Ideogram 4 model-card: 48 steps, guidance 7 (its schedule tapers the last 3 steps to 3.0;
     # the loader keeps that taper when the request matches these defaults exactly).
     ("ideogram", 48, 7.0),
diff --git a/studio/backend/core/inference/diffusion_hidream.py b/studio/backend/core/inference/diffusion_hidream.py
new file mode 100644
index 0000000000..82ac27f7d5
--- /dev/null
+++ b/studio/backend/core/inference/diffusion_hidream.py
@@ -0,0 +1,47 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""HiDream-I1 Llama text-encoder assembly.
+
+The HiDream-ai/HiDream-I1-* repos name ``text_encoder_4`` (LlamaForCausalLM) and
+``tokenizer_4`` in their model_index but do NOT ship the weights: the official example
+loads meta-llama/Meta-Llama-3.1-8B-Instruct separately and passes both components into
+``HiDreamImagePipeline.from_pretrained``. That upstream repo is Hub-gated (manual
+approval), so Studio loads the open unsloth mirror instead -- byte-identical weights,
+no license wall at load time, and the unsloth org is already inside the loader's
+non-GGUF trust gate. ``output_hidden_states=True`` matches the official example: the
+pipeline's prompt encoder consumes the Llama hidden states, not the logits.
+"""
+
+from __future__ import annotations
+
+from typing import Any, Optional
+
+from loggers import get_logger
+
+logger = get_logger(__name__)
+
+HIDREAM_FAMILY_NAME = "hidream-i1"
+
+# Open mirror of the gated meta-llama/Meta-Llama-3.1-8B-Instruct the pipeline expects.
+HIDREAM_LLAMA_REPO = "unsloth/Meta-Llama-3.1-8B-Instruct"
+
+
+def hidream_te4_kwargs(dtype: Any, hf_token: Optional[str] = None) -> dict[str, Any]:
+    """``{text_encoder_4, tokenizer_4}`` kwargs for a HiDream pipeline ``from_pretrained``.
+
+    Loaded eagerly (~16 GB bf16) before the pipeline call so a failure surfaces as a
+    clear error instead of a half-built pipeline."""
+    import torch  # noqa: F401 -- dtype values are torch dtypes; import keeps parity with callers
+    from transformers import AutoTokenizer, LlamaForCausalLM
+
+    logger.info("diffusion.hidream: loading Llama TE4 from %s", HIDREAM_LLAMA_REPO)
+    tokenizer_4 = AutoTokenizer.from_pretrained(HIDREAM_LLAMA_REPO, token = hf_token)
+    text_encoder_4 = LlamaForCausalLM.from_pretrained(
+        HIDREAM_LLAMA_REPO,
+        output_hidden_states = True,
+        output_attentions = True,
+        torch_dtype = dtype,
+        token = hf_token,
+    )
+    return {"text_encoder_4": text_encoder_4, "tokenizer_4": tokenizer_4}
diff --git a/studio/backend/tests/test_diffusion_more_families.py b/studio/backend/tests/test_diffusion_more_families.py
index f59988b7ac..afd0f3df7c 100644
--- a/studio/backend/tests/test_diffusion_more_families.py
+++ b/studio/backend/tests/test_diffusion_more_families.py
@@ -202,6 +202,62 @@ def test_hunyuanimage21_bf16_component_table_present():
     assert vae_gb <= 1.0
 
 
+# ── hidream-i1 family ────────────────────────────────────────────────────────
+@pytest.mark.parametrize(
+    "repo_id",
+    [
+        "HiDream-ai/HiDream-I1-Full",
+        "HiDream-ai/HiDream-I1-Dev",
+        "HiDream-ai/HiDream-I1-Fast",
+    ],
+)
+def test_detect_family_hidream_repos(repo_id):
+    # One family covers all three variants (same 17B MoE arch + 4-TE stack).
+    fam = detect_family(repo_id)
+    assert fam is not None and fam.name == "hidream-i1"
+    assert fam.pipeline_class == "HiDreamImagePipeline"
+    assert fam.transformer_class == "HiDreamImageTransformer2DModel"
+    assert fam.base_repo == "HiDream-ai/HiDream-I1-Full"
+    # Published bf16-only upstream; the fp16 fallback stays off.
+    assert fam.fp16_incompatible is True
+
+
+def test_hidream_override_and_trust():
+    assert detect_family("x", override = "hidream-i1").name == "hidream-i1"
+    assert detect_family("x", override = "hidream").name == "hidream-i1"
+    # The three official repos load via from_pretrained -> allowlisted; the Llama TE4
+    # comes from the unsloth mirror, which the org prefix already trusts.
+    for rid in (
+        "HiDream-ai/HiDream-I1-Full",
+        "HiDream-ai/HiDream-I1-Dev",
+        "HiDream-ai/HiDream-I1-Fast",
+    ):
+        assert _is_trusted_diffusion_repo(rid)
+    assert not _is_trusted_diffusion_repo("HiDream-ai/some-future-repo")
+    assert _is_trusted_diffusion_repo("unsloth/Meta-Llama-3.1-8B-Instruct")
+
+
+def test_hidream_generation_defaults():
+    # Upstream inference.py: Full 50 steps / guidance 5; Dev and Fast are distilled and
+    # run guidance-free at 28 / 16 steps. The specific keys must beat the generic
+    # "hidream" (which also appears in the owner segment of every variant id).
+    assert default_generation_params("HiDream-ai/HiDream-I1-Full") == (50, 5.0)
+    assert default_generation_params("HiDream-ai/HiDream-I1-Dev") == (28, 0.0)
+    assert default_generation_params("HiDream-ai/HiDream-I1-Fast") == (16, 0.0)
+
+
+def test_hidream_bf16_component_table_present():
+    fam = detect_family("HiDream-ai/HiDream-I1-Full")
+    sizes = family_bf16_components_gb(fam)
+    assert sizes is not None
+    transformer_gb, encoders_gb, vae_gb = sizes
+    # 17B MoE DiT 34.2 GB; TEs = CLIP-L 0.5 + CLIP-G 2.8 + T5-XXL 9.5 from the repo plus
+    # the ~16 GB Llama TE4 assembled from the mirror -> ~28.8 GB.
+    assert 32.0 <= transformer_gb <= 37.0
+    assert 26.0 <= encoders_gb <= 32.0
+    assert vae_gb <= 0.5
+
+
 def test_ideogram4_generation_defaults():
     # Model-card settings: 48 steps, guidance 7 (the backend keeps the pipeline's
     # recommended tapered schedule when the request matches exactly).