diff --git a/src/maxtext/configs/models/kimi-k3.yml b/src/maxtext/configs/models/kimi-k3.yml
index e2794e7049..b21fc4289e 100644
--- a/src/maxtext/configs/models/kimi-k3.yml
+++ b/src/maxtext/configs/models/kimi-k3.yml
@@ -48,7 +48,7 @@ attn_res_block_size: 12
 
 # MLP and MoE Specifications
 first_num_dense_layers: 1
-mlp_activations: ["situ", "linear"]
+mlp_activations: ["situ_gate", "situ_linear"]
 # `base_mlp_dim` is the layer-0 dense MLP width (HF `intermediate_size`); the MoE
 # shared expert uses `shared_intermediate_size` = moe_intermediate_size * num_shared_experts.
 base_mlp_dim: 33792
diff --git a/src/maxtext/layers/linears.py b/src/maxtext/layers/linears.py
index b6bcdd6a70..cd02ff90a5 100644
--- a/src/maxtext/layers/linears.py
+++ b/src/maxtext/layers/linears.py
@@ -44,6 +44,59 @@
 from maxtext.utils.sharding import truncate_out_sharding
 
 
+# SiTU betas used by Kimi-K3; they come from the HF config keys `activation_situ_beta`
+# and `activation_situ_linear_beta` (HF `SituAndMul`) and are not configurable here.
+# They are the explicit Kimi-K3 preset used as defaults below and by the registered
+# `"situ_gate"` / `"situ_linear"` activation names (HF `KimiLinearConfig` itself defaults
+# to `beta=1.0, linear_beta=None`, which is not what Kimi-K3 ships with).
+KIMI_K3_SITU_BETA = 4.0
+KIMI_K3_SITU_LINEAR_BETA = 25.0
+
+
+def _check_situ_beta(name: str, beta: float) -> None:
+  """Rejects non-positive SiTU betas (`beta=0` would otherwise produce `0 / 0 = NaN`)."""
+  if not beta > 0:
+    raise ValueError(f"SiTU {name} must be > 0, got {beta}.")
+
+
+def situ_gate(x: Array, beta: float = KIMI_K3_SITU_BETA) -> Array:
+  """SiTU gate activation: beta * tanh(x / beta) * sigmoid(x), computed in fp32."""
+  _check_situ_beta("beta", beta)
+  x32 = x.astype(jnp.float32)
+  return (beta * jnp.tanh(x32 / beta) * jax.nn.sigmoid(x32)).astype(x.dtype)
+
+
+def situ_linear(x: Array, beta: float = KIMI_K3_SITU_LINEAR_BETA) -> Array:
+  """SiTU soft-capped linear (up-branch) activation: beta * tanh(x / beta), computed in fp32."""
+  _check_situ_beta("linear_beta", beta)
+  x32 = x.astype(jnp.float32)
+  return (beta * jnp.tanh(x32 / beta)).astype(x.dtype)
+
+
+def situ_and_mul(
+    gate: Array,
+    up: Array,
+    *,
+    gate_beta: float = KIMI_K3_SITU_BETA,
+    linear_beta: float | None = KIMI_K3_SITU_LINEAR_BETA,
+    dtype: DType | None = None,
+) -> Array:
+  """Fused SiTU GLU: situ_gate(gate) * situ_linear(up), matching HF `SituAndMul`.
+
+  Both branches and their product are computed in fp32 and cast to `dtype` (default:
+  `gate.dtype`) once at the end, so bf16 callers round once instead of three times.
+  `linear_beta=None` leaves `up` untransformed, as in HF.
+  """
+  _check_situ_beta("beta", gate_beta)
+  g32 = gate.astype(jnp.float32)
+  u32 = up.astype(jnp.float32)
+  situ_a = gate_beta * jnp.tanh(g32 / gate_beta) * jax.nn.sigmoid(g32)
+  if linear_beta is not None:
+    _check_situ_beta("linear_beta", linear_beta)
+    u32 = linear_beta * jnp.tanh(u32 / linear_beta)
+  return (situ_a * u32).astype(gate.dtype if dtype is None else dtype)
+
+
 def _convert_to_activation_function(fn_or_string: str | Callable[..., Any]) -> Callable[..., Any]:
   """Convert a string to an activation function."""
   if fn_or_string == "linear":
@@ -51,6 +104,11 @@ def _convert_to_activation_function(fn_or_string: str | Callable[..., Any]) -> C
   elif fn_or_string == "sqrtsoftplus":
     # Custom activation function used by DeepSeek V4 Top-K MoE router
     return lambda x: jnp.sqrt(jax.nn.softplus(x))
+  # SiTU activations used by the Kimi-K3 MLP/experts (gate and up branches)
+  elif fn_or_string == "situ_gate":
+    return situ_gate
+  elif fn_or_string == "situ_linear":
+    return situ_linear
   elif isinstance(fn_or_string, str):
     return getattr(jax.nn, fn_or_string)
   elif callable(fn_or_string):
@@ -809,6 +867,7 @@ def __call__(
     # Iterate over specified MLP input activation functions.
     # e.g. ('relu',) or ('gelu', 'linear') for gated-gelu.
     activations = []
+    is_situ_glu = tuple(self.activations) == ("situ_gate", "situ_linear")
     if cfg.fused_mlp:
       x = self.wi(inputs, out_sharding=intermediate_sharding)
 
@@ -818,8 +877,8 @@ def __call__(
 
       x = checkpoint_name(x, "mlpwi")
       for idx, act_fn in enumerate(self.activations):
-        y = _convert_to_activation_function(act_fn)(self._clip_glu_input(idx, x[:, :, idx, ...]))
-        activations.append(y)
+        y = self._clip_glu_input(idx, x[:, :, idx, ...])
+        activations.append(y if is_situ_glu else _convert_to_activation_function(act_fn)(y))
     else:
       for idx, act_fn in enumerate(self.activations):
         dense_name = "wi" if len(self.activations) == 1 else f"wi_{idx}"
@@ -828,11 +887,15 @@ def __call__(
         x = checkpoint_name(x, "mlp" + dense_name)
         if cfg.activations_in_float32:
           x = x.astype(jnp.float32)
-        x = _convert_to_activation_function(act_fn)(self._clip_glu_input(idx, x))
-        activations.append(x)
+        x = self._clip_glu_input(idx, x)
+        activations.append(x if is_situ_glu else _convert_to_activation_function(act_fn)(x))
 
-    # Take elementwise product of above intermediate activations.
-    x = functools.reduce(operator.mul, activations).astype(self.dtype)
+    if is_situ_glu:
+      # SiTU GLU: compute gate * up in fp32 and round once, matching HF `SituAndMul`.
+      x = situ_and_mul(activations[0], activations[1], dtype=self.dtype)
+    else:
+      # Take elementwise product of above intermediate activations.
+      x = functools.reduce(operator.mul, activations).astype(self.dtype)
     # Apply dropout and final dense output projection.
     x = self.dropout(x, deterministic=deterministic)  # Broadcast along length.
     x = self._maybe_shard_with_logical(x, self.intermediate_logical)
diff --git a/tests/unit/kimi_k3_situ_activation_test.py b/tests/unit/kimi_k3_situ_activation_test.py
new file mode 100644
index 0000000000..6b1bc00303
--- /dev/null
+++ b/tests/unit/kimi_k3_situ_activation_test.py
@@ -0,0 +1,247 @@
+# Copyright 2026 Google LLC
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""Unit tests for the Kimi-K3 SiTU activation functions (situ_gate / situ_linear)."""
+
+import ast
+import json
+import os
+import sys
+import unittest
+from unittest import mock
+
+from flax import nnx
+import jax
+import jax.numpy as jnp
+from jax.test_util import check_grads
+import numpy as np
+
+from maxtext.configs import pyconfig
+from maxtext.layers import linears
+from maxtext.layers.linears import situ_and_mul, situ_gate, situ_linear, _convert_to_activation_function
+from maxtext.utils import maxtext_utils
+from tests.utils.kimi_k3_reference import KIMI_K3_REFERENCE_DIR, requires_kimi_k3_reference
+from tests.utils.test_helpers import get_test_config_path
+
+try:
+  import torch  # pylint: disable=import-error
+except ImportError:  # pragma: no cover - torch is optional for the hermetic tests.
+  torch = None
+
+
+def _load_reference_situ_and_mul():
+  """Loads HF `SituAndMul` from the reference without importing `modeling_kimi_linear`.
+
+  Importing the full reference module needs `fla` and a recent `transformers`; extracting just
+  the class source keeps this test free of sys.modules / sys.path side effects.
+  """
+  path = os.path.join(KIMI_K3_REFERENCE_DIR, "modeling_kimi_linear.py")
+  with open(path, encoding="utf-8") as f:
+    tree = ast.parse(f.read())
+  cls = next(n for n in tree.body if isinstance(n, ast.ClassDef) and n.name == "SituAndMul")
+  namespace = {"torch": torch, "nn": torch.nn}
+  exec(compile(ast.Module(body=[cls], type_ignores=[]), path, "exec"), namespace)  # pylint: disable=exec-used
+  return namespace["SituAndMul"]
+
+
+def _np_situ_and_mul(gate, up, cotangent, beta=4.0, linear_beta=25.0):
+  """Hermetic float64 NumPy oracle for HF `SituAndMul`: returns (y, d_gate, d_up) for the VJP."""
+  g = np.asarray(gate, np.float64)
+  u = np.asarray(up, np.float64)
+  ct = np.asarray(cotangent, np.float64)
+  sig = 1.0 / (1.0 + np.exp(-g))
+  tg = np.tanh(g / beta)
+  situ_a = beta * tg * sig
+  d_situ_a = (1.0 - tg**2) * sig + beta * tg * sig * (1.0 - sig)
+  tu = np.tanh(u / linear_beta)
+  lin = linear_beta * tu
+  d_lin = 1.0 - tu**2
+  return situ_a * lin, ct * d_situ_a * lin, ct * situ_a * d_lin
+
+
+def _situ_inputs(dtype, shape=(8, 128)):
+  """Gate/up/cotangent inputs spanning the soft-cap regime ([-30, 30] and beyond)."""
+  k1, k2, k3 = jax.random.split(jax.random.PRNGKey(0), 3)
+  gate = (jax.random.normal(k1, shape, jnp.float32) * 15.0).astype(dtype)
+  up = (jax.random.normal(k2, shape, jnp.float32) * 40.0).astype(dtype)
+  ct = jax.random.normal(k3, shape, jnp.float32).astype(dtype)
+  return gate, up, ct
+
+
+# Tolerances vs the float64 oracle. bf16 results are rounded once (half a bf16 ulp ~ 2e-3 relative). The fp32
+# atol covers fp32 cancellation in `1 - tanh(x / beta)**2` (abs error ~1e-5) in the gate gradient.
+_SITU_TOLS = {jnp.float32: {"rtol": 1e-5, "atol": 5e-5}, jnp.bfloat16: {"rtol": 4e-3, "atol": 1e-4}}
+
+
+class KimiK3SituActivationTest(unittest.TestCase):
+  """Tests validating SiTU activation values, properties and reference parity."""
+
+  def test_registration(self):
+    self.assertIs(_convert_to_activation_function("situ_gate"), situ_gate)
+    self.assertIs(_convert_to_activation_function("situ_linear"), situ_linear)
+
+  def test_golden_values(self):
+    """Hard-coded values computed independently (Python math) for beta=4 / linear_beta=25."""
+    x_gate = jnp.array([0.0, 1.0, -2.0, 10.0], dtype=jnp.float32)
+    np.testing.assert_allclose(
+        situ_gate(x_gate), [0.0, 0.7161995568672704, -0.22034286184779028, 3.9462780318602517], rtol=1e-6, atol=1e-7
+    )
+    x_lin = jnp.array([0.0, 1.0, 100.0, -50.0], dtype=jnp.float32)
+    np.testing.assert_allclose(
+        situ_linear(x_lin), [0.0, 0.9994670077790893, 24.983232493476677, -24.100689501895424], rtol=1e-6, atol=1e-7
+    )
+
+  def test_asymptotes_and_gradients(self):
+    big = jnp.array([1e4, -1e4], dtype=jnp.float32)
+    np.testing.assert_allclose(situ_linear(big), [25.0, -25.0], rtol=1e-6)
+    np.testing.assert_allclose(situ_gate(big), [4.0, 0.0], atol=1e-6)
+    small = jnp.linspace(-0.1, 0.1, 11, dtype=jnp.float32)
+    np.testing.assert_allclose(situ_linear(small), small, atol=1e-5)
+    for fn in (situ_gate, situ_linear):
+      grads = jax.vmap(jax.grad(fn))(jnp.array([-1e3, -30.0, 0.0, 30.0, 1e3], dtype=jnp.float32))
+      self.assertTrue(bool(jnp.all(jnp.isfinite(grads))))
+
+  def test_bf16_preserves_dtype_and_matches_fp32(self):
+    x = jax.random.normal(jax.random.PRNGKey(0), (256,), dtype=jnp.float32) * 20.0
+    x_bf16 = x.astype(jnp.bfloat16)
+    for fn in (situ_gate, situ_linear):
+      y = fn(x_bf16)
+      self.assertEqual(y.dtype, jnp.bfloat16)
+      # fp32 compute + single cast back: result equals the fp32 result rounded to bf16.
+      expected = fn(x_bf16.astype(jnp.float32)).astype(jnp.bfloat16)
+      np.testing.assert_array_equal(np.asarray(y, np.float32), np.asarray(expected, np.float32))
+
+  def test_situ_and_mul_forward_and_vjp_match_numpy_oracle(self):
+    """Hermetic (CI) check of the combined gate-times-up math and its VJP, fp32 and bf16."""
+    for dtype in (jnp.float32, jnp.bfloat16):
+      with self.subTest(dtype=dtype.__name__):
+        gate, up, ct = _situ_inputs(dtype)
+        y, vjp_fn = jax.vjp(situ_and_mul, gate, up)
+        d_gate, d_up = vjp_fn(ct)
+        for arr in (y, d_gate, d_up):
+          self.assertEqual(arr.dtype, dtype)
+        y_ref, d_gate_ref, d_up_ref = _np_situ_and_mul(
+            np.asarray(gate, np.float32), np.asarray(up, np.float32), np.asarray(ct, np.float32)
+        )
+        tol = _SITU_TOLS[dtype]
+        np.testing.assert_allclose(np.asarray(y, np.float32), y_ref, **tol)
+        np.testing.assert_allclose(np.asarray(d_gate, np.float32), d_gate_ref, **tol)
+        np.testing.assert_allclose(np.asarray(d_up, np.float32), d_up_ref, **tol)
+
+  def test_composed_fp32_matches_numpy_oracle(self):
+    """situ_gate(gate) * situ_linear(up) in fp32 (and its VJP) is the same math as the fused helper."""
+    gate, up, ct = _situ_inputs(jnp.float32)
+    y, vjp_fn = jax.vjp(lambda g, u: situ_gate(g) * situ_linear(u), gate, up)
+    d_gate, d_up = vjp_fn(ct)
+    y_ref, d_gate_ref, d_up_ref = _np_situ_and_mul(gate, up, ct)
+    for got, want in ((y, y_ref), (d_gate, d_gate_ref), (d_up, d_up_ref)):
+      np.testing.assert_allclose(np.asarray(got), want, **_SITU_TOLS[jnp.float32])
+
+  def test_tolerance_rejects_perturbed_betas(self):
+    """Negative control: a +1% beta error must fail the tolerances used above (incl. bf16)."""
+    for dtype in (jnp.float32, jnp.bfloat16):
+      gate, up, ct = _situ_inputs(dtype)
+      y_ref, _, _ = _np_situ_and_mul(np.asarray(gate, np.float32), np.asarray(up, np.float32), np.asarray(ct, np.float32))
+      for kwargs in ({"gate_beta": 4.04}, {"linear_beta": 25.25}):
+        with self.subTest(dtype=dtype.__name__, **kwargs):
+          y = situ_and_mul(gate, up, **kwargs)
+          with self.assertRaises(AssertionError):
+            np.testing.assert_allclose(np.asarray(y, np.float32), y_ref, **_SITU_TOLS[dtype])
+
+  def test_situ_and_mul_bf16_rounds_once(self):
+    """bf16 output equals the fp32 result cast to bf16 once (HF `SituAndMul` semantics)."""
+    gate, up, _ = _situ_inputs(jnp.bfloat16)
+    y = situ_and_mul(gate, up)
+    expected = situ_and_mul(gate.astype(jnp.float32), up.astype(jnp.float32)).astype(jnp.bfloat16)
+    np.testing.assert_array_equal(np.asarray(y, np.float32), np.asarray(expected, np.float32))
+    self.assertEqual(situ_and_mul(gate, up, dtype=jnp.float32).dtype, jnp.float32)
+
+  def test_situ_and_mul_linear_beta_none(self):
+    """`linear_beta=None` leaves the up branch untransformed, as in HF."""
+    gate, up, _ = _situ_inputs(jnp.float32)
+    np.testing.assert_allclose(situ_and_mul(gate, up, linear_beta=None), situ_gate(gate) * up, rtol=1e-6, atol=1e-6)
+
+  def test_check_grads(self):
+    """Numerical-vs-autodiff gradient check (first and second order) in fp32."""
+    x = jnp.array([-12.0, -3.0, -0.5, 0.0, 0.7, 2.5, 9.0, 40.0], jnp.float32)
+    check_grads(situ_gate, (x,), order=2, modes=["fwd", "rev"], atol=1e-2, rtol=1e-2, eps=1e-2)
+    check_grads(situ_linear, (x,), order=2, modes=["fwd", "rev"], atol=1e-2, rtol=1e-2, eps=1e-2)
+    check_grads(situ_and_mul, (x, x[::-1]), order=1, modes=["fwd", "rev"], atol=1e-2, rtol=1e-2, eps=1e-2)
+
+  def test_nonpositive_beta_rejected(self):
+    x = jnp.ones((2,), jnp.float32)
+    for beta in (0.0, -1.0):
+      with self.assertRaises(ValueError):
+        situ_gate(x, beta=beta)
+      with self.assertRaises(ValueError):
+        situ_linear(x, beta=beta)
+      with self.assertRaises(ValueError):
+        situ_and_mul(x, x, gate_beta=beta)
+      with self.assertRaises(ValueError):
+        situ_and_mul(x, x, linear_beta=beta)
+
+  def test_mlp_block_uses_fused_situ_and_mul(self):
+    """MlpBlock with the SiTU pair multiplies in fp32 and rounds once (fused and unfused wi)."""
+    for fused_mlp in (False, True):
+      with self.subTest(fused_mlp=fused_mlp):
+        cfg = pyconfig.initialize(
+            [sys.argv[0], get_test_config_path()],
+            per_device_batch_size=1.0,
+            run_name="test",
+            enable_checkpointing=False,
+            max_target_length=128,
+            fused_mlp=fused_mlp,
+        )
+        mesh = jax.sharding.Mesh(maxtext_utils.create_device_mesh(cfg), cfg.mesh_axes)
+        layer = linears.MlpBlock(
+            config=cfg,
+            mesh=mesh,
+            in_features=16,
+            intermediate_dim=32,
+            activations=("situ_gate", "situ_linear"),
+            intermediate_dropout_rate=0.0,
+            dtype=jnp.bfloat16,
+            rngs=nnx.Rngs(params=0, dropout=1),
+        )
+        inputs = jax.random.normal(jax.random.PRNGKey(1), (2, 3, 16), jnp.float32).astype(jnp.bfloat16) * 8
+        with mock.patch.object(linears, "situ_and_mul", wraps=linears.situ_and_mul) as spy:
+          out = layer(inputs, deterministic=True)
+        spy.assert_called_once()
+        self.assertEqual(out.shape, (2, 3, 16))
+        self.assertTrue(bool(jnp.all(jnp.isfinite(out))))
+
+  @unittest.skipIf(torch is None, "torch is not installed")
+  @requires_kimi_k3_reference
+  def test_parity_against_reference(self):
+    """Numerical parity against HF SituAndMul, covering the soft-cap regime and bf16."""
+    with open(os.path.join(KIMI_K3_REFERENCE_DIR, "config.json"), encoding="utf-8") as f:
+      hf_config = json.load(f)
+    hf_config = hf_config.get("text_config", hf_config)
+    beta = hf_config["activation_situ_beta"]
+    linear_beta = hf_config["activation_situ_linear_beta"]
+    self.assertEqual((beta, linear_beta), (linears.KIMI_K3_SITU_BETA, linears.KIMI_K3_SITU_LINEAR_BETA))
+    ref_situ_and_mul = _load_reference_situ_and_mul()(beta=beta, linear_beta=linear_beta)
+    torch.manual_seed(42)
+    x_pt = torch.randn(4, 16, 64, dtype=torch.float32) * 20.0
+    for dtype_pt, dtype_jax in ((torch.float32, jnp.float32), (torch.bfloat16, jnp.bfloat16)):
+      x = x_pt.to(dtype_pt)
+      y_pt = ref_situ_and_mul(x).float().numpy()
+      x_jax = jnp.asarray(x.float().numpy()).astype(dtype_jax)
+      y_jax = situ_and_mul(x_jax[..., :32], x_jax[..., 32:], gate_beta=beta, linear_beta=linear_beta)
+      self.assertEqual(y_jax.dtype, dtype_jax)
+      np.testing.assert_allclose(np.asarray(y_jax, np.float32), y_pt, **_SITU_TOLS[dtype_jax])
+
+
+if __name__ == "__main__":
+  unittest.main()