diff --git a/doc/APIreference/APIglobals.rst b/doc/APIreference/APIglobals.rst
index f1ad2c5b..2d08ce08 100644
--- a/doc/APIreference/APIglobals.rst
+++ b/doc/APIreference/APIglobals.rst
@@ -534,6 +534,14 @@ shown in the table below. Their names are in the format ``mjKEY_XXX``. They corr
 Macros
 ^^^^^^
 
+
+.. _mjUSESINGLE:
+
+mjUSESINGLE
+~~~~~~~~~~~
+
+Compile-time flag, see :ref:`mjtNum`.
+
 .. _mjDISABLED:
 
 mjDISABLED
diff --git a/doc/APIreference/APItypes.rst b/doc/APIreference/APItypes.rst
index 5ee07e6a..e5ff73ff 100644
--- a/doc/APIreference/APItypes.rst
+++ b/doc/APIreference/APItypes.rst
@@ -54,24 +54,28 @@ The two types below are defined in `mjtnum.h <https://github.com/google-deepmind
 mjtNum
 ^^^^^^
 
-This is the floating-point type used throughout the simulator. If the symbol ``mjUSEDOUBLE`` is defined in
-``mjmodel.h``, this type is defined as ``double``, otherwise it is defined as ``float``. Currently only the
-double-precision version of MuJoCo is distributed, although the entire code base works with single-precision as well.
-We may release the single-precision version in the future for efficiency reasons, but the double-precision version
-will always be available. Thus it is safe to write user code assuming double precision. However, our preference is to
-write code that works with either single or double precision. To this end we provide math utility functions that are
-always defined with the correct floating-point type.
+This is the floating-point type used throughout the simulator. When using the default build configuration, ``mjtNum`` is
+defined as ``double``. If the symbol ``mjUSESINGLE`` is defined, ``mjtNum`` is defined as ``float``.
 
-Note that changing ``mjUSEDOUBLE`` in ``mjtnum.h`` will not change how the library was compiled, and instead will
+Currently only the double-precision version of MuJoCo is distributed, although the entire code base works with
+single-precision as well. We may release the single-precision version in the future, but the
+double-precision version will always be available. Thus it is safe to write user code assuming double precision.
+However, our preference is to write code that works with either single or double precision. To this end we provide math
+utility functions that are always defined with the correct floating-point type.
+
+Note that changing ``mjUSESINGLE`` in ``mjtnum.h`` will not change how the library was compiled, and instead will
 result in numerous link errors. In general, the header files distributed with precompiled MuJoCo should never be
 changed by the user.
 
 .. code-block:: C
 
-   #ifdef mjUSEDOUBLE
-       typedef double mjtNum;
+   // floating point data type and minval
+   #ifndef mjUSESINGLE
+     typedef double mjtNum;
+     #define mjMINVAL    1E-15       // minimum value in any denominator
    #else
-       typedef float mjtNum;
+     typedef float mjtNum;
+     #define mjMINVAL    1E-15f
    #endif
 
 
diff --git a/doc/changelog.rst b/doc/changelog.rst
index ff007373..cdef866c 100644
--- a/doc/changelog.rst
+++ b/doc/changelog.rst
@@ -25,12 +25,18 @@ General
 5. Added support for ``ball`` joints in the URDF parser ("spherical" in URDF).
 6. Deprecated :ref:`mju_rotVecMat` and :ref:`mju_rotVecMatT` in favor of :ref:`mju_mulMatVec3` and
    :ref:`mju_mulMatTVec3`. These functions names and argument ordering are more consistent with the rest of the API.
+7. Replaced ``mjUSEDOUBLE`` which was previously hard-coded in
+   `mjtnum.h <https://github.com/google-deepmind/mujoco/blob/main/include/mujoco/mjtnum.h>`__
+   with the build-time flag ``mjUSESINGLE``. If this symbol is not defined, MuJoCo will use double-precision floating
+   point, as usual. If ``mjUSESINGLE`` is defined, MuJoCo will use single-precision floating point. See :ref:`mjtNum`.
+
+   Relatedly, fixed various type errors that prevented building with single-precision.
 
 MJX
 ~~~
-7. Added support for :ref:`elliptic friction cones<option-cone>`.
-8. Fixed a bug that resulted in less-optimal linesearch solutions for some difficult constraint settings.
-9. Fixed a bug in the Newton solver that sometimes resulted in less-optimal gradients.
+8. Added support for :ref:`elliptic friction cones<option-cone>`.
+9. Fixed a bug that resulted in less-optimal linesearch solutions for some difficult constraint settings.
+10. Fixed a bug in the Newton solver that sometimes resulted in less-optimal gradients.
 
 
 .. youtube:: P83tKA1iz2Y
@@ -39,8 +45,8 @@ MJX
 
 Simulate
 ^^^^^^^^
-10. Added improved tutorial video.
-11. Improved the Brownian noise generator.
+11. Added improved tutorial video.
+12. Improved the Brownian noise generator.
 
 |br| |br| |br| |br|
 
diff --git a/include/mujoco/mjtnum.h b/include/mujoco/mjtnum.h
index b9d78ea1..e50b5ff5 100644
--- a/include/mujoco/mjtnum.h
+++ b/include/mujoco/mjtnum.h
@@ -17,12 +17,8 @@
 
 //---------------------------------- floating-point definition -------------------------------------
 
-// compile-time configuration options
-#define mjUSEDOUBLE               // single or double precision for mjtNum
-
-
 // floating point data type and minval
-#ifdef mjUSEDOUBLE
+#ifndef mjUSESINGLE
   typedef double mjtNum;
   #define mjMINVAL    1E-15       // minimum value in any denominator
 #else
diff --git a/include/mujoco/mujoco.h b/include/mujoco/mujoco.h
index 95893542..1c85099f 100644
--- a/include/mujoco/mujoco.h
+++ b/include/mujoco/mujoco.h
@@ -893,7 +893,7 @@ MJAPI int mjs_isWarning(mjSpec* s);
 
 //---------------------------------- Standard math -------------------------------------------------
 
-#ifdef mjUSEDOUBLE
+#if !defined(mjUSESINGLE)
   #define mju_sqrt    sqrt
   #define mju_exp     exp
   #define mju_sin     sin
diff --git a/sample/testspeed.cc b/sample/testspeed.cc
index 66dabf3a..09757888 100644
--- a/sample/testspeed.cc
+++ b/sample/testspeed.cc
@@ -34,18 +34,17 @@ mjData* d[maxthread];
 // per-thread statistics
 int contacts[maxthread];
 int constraints[maxthread];
-double simtime[maxthread];
-
+mjtNum simtime[maxthread];
 
 // timer
-std::chrono::steady_clock::time_point tm_start;
 mjtNum gettm(void) {
-  std::chrono::duration<double, std::micro> elapsed;
-  elapsed = std::chrono::steady_clock::now() - tm_start;
+  using std::chrono::steady_clock;
+  using Microseconds = std::chrono::duration<double, std::micro>;
+  static steady_clock::time_point tm_start = steady_clock::now();
+  auto elapsed = Microseconds(steady_clock::now() - tm_start);
   return elapsed.count();
 }
 
-
 // deallocate and print message
 int finish(const char* msg = NULL, mjModel* m = NULL) {
   // deallocate model
@@ -87,7 +86,7 @@ void simulate(int id, int nstep, mjtNum* ctrl) {
   constraints[id] = 0;
 
   // run and time
-  double start = gettm();
+  mjtNum start = gettm();
   for (int i=0; i < nstep; i++) {
     // inject pseudo-random control noise
     mju_copy(d[id]->ctrl, ctrl + i*m->nu, m->nu);
@@ -126,7 +125,7 @@ int main(int argc, char** argv) {
   // read arguments
   int nstep = 10000, nthread = 0, npoolthread = 0;
   // inject small noise by default, to avoid fixed contact state
-  mjtNum ctrlnoise = 0.01;
+  double ctrlnoise = 0.01;
   if (argc > 2 && (std::sscanf(argv[2], "%d", &nstep) != 1 || nstep <= 0)) {
     return finish("Invalid nstep argument");
   }
@@ -149,7 +148,7 @@ int main(int argc, char** argv) {
 
   // get filename, determine file type
   std::string filename(argv[1]);
-  bool binary = (filename.find(".mjb") != std::string::npos);
+  bool binary = (filename.find(".mjb") != std::string::npos);  // NOLINT
 
   // load model
   char error[1000] = "Could not load binary model";
@@ -191,8 +190,11 @@ int main(int argc, char** argv) {
               nstep,
               nthread > 1 ? " per thread" : "",
               m->opt.timestep);
+  if (sizeof(mjtNum) == 4) {
+    std::printf(", using single-precision");
+  }
   if (npoolthread > 1) {
-    std::printf(", using %d threads for engine-internal threadpool", npoolthread);
+    std::printf(", using %d threads", npoolthread);
   }
   std::printf("...\n\n");
 
diff --git a/simulate/main.cc b/simulate/main.cc
index 30139f8f..397a83c9 100644
--- a/simulate/main.cc
+++ b/simulate/main.cc
@@ -12,6 +12,7 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.
 
+#include <cerrno>
 #include <chrono>
 #include <cstdint>
 #include <cstdio>
@@ -334,7 +335,7 @@ void PhysicsLoop(mj::Simulate& sim) {
 
           // misalignment condition: distance from target sim time is bigger than syncmisalign
           bool misaligned =
-              mju_abs(Seconds(elapsedCPU).count()/slowdown - elapsedSim) > syncMisalign;
+              std::abs(Seconds(elapsedCPU).count()/slowdown - elapsedSim) > syncMisalign;
 
           // out-of-sync (for any reason): reset sync times, step
           if (elapsedSim < 0 || elapsedCPU.count() < 0 || syncCPU.time_since_epoch().count() == 0 ||
diff --git a/simulate/simulate.cc b/simulate/simulate.cc
index a86c7343..a424e310 100644
--- a/simulate/simulate.cc
+++ b/simulate/simulate.cc
@@ -1165,7 +1165,9 @@ void CopyPose(mj::Simulate* sim, const mjModel* m, const mjData* d) {
 
 // millisecond timer, for MuJoCo built-in profiler
 mjtNum Timer() {
-  return Milliseconds(mj::Simulate::Clock::now().time_since_epoch()).count();
+  static auto start = mj::Simulate::Clock::now();
+  auto elapsed = Milliseconds(mj::Simulate::Clock::now() - start);
+  return elapsed.count();
 }
 
 // clear all times
diff --git a/src/engine/engine_core_constraint.c b/src/engine/engine_core_constraint.c
index 6e47d384..e17c2bad 100644
--- a/src/engine/engine_core_constraint.c
+++ b/src/engine/engine_core_constraint.c
@@ -35,9 +35,9 @@
 #endif
 
 #ifdef mjUSEPLATFORMSIMD
-  #if defined(__AVX__) && defined(mjUSEDOUBLE)
+  #if defined(__AVX__) && !defined(mjUSESINGLE)
     #define mjUSEAVX
-  #endif  // defined(__AVX__) && defined(mjUSEDOUBLE)
+  #endif  // defined(__AVX__) && !defined(mjUSESINGLE)
 #endif  // mjUSEPLATFORMSIMD
 
 
diff --git a/src/engine/engine_support.c b/src/engine/engine_support.c
index 526a1334..6740f2ae 100644
--- a/src/engine/engine_support.c
+++ b/src/engine/engine_support.c
@@ -31,7 +31,7 @@
 #include "engine/engine_util_spatial.h"
 
 #ifdef mjUSEPLATFORMSIMD
-  #if defined(__AVX__) && defined(mjUSEDOUBLE)
+  #if defined(__AVX__) && !defined(mjUSESINGLE)
     #define mjUSEAVX
     #include "immintrin.h"
   #endif
diff --git a/src/engine/engine_util_blas.c b/src/engine/engine_util_blas.c
index bb920b86..77ed4194 100644
--- a/src/engine/engine_util_blas.c
+++ b/src/engine/engine_util_blas.c
@@ -19,7 +19,7 @@
 #include <mujoco/mjtnum.h>
 
 #ifdef mjUSEPLATFORMSIMD
-  #if defined(__AVX__) && defined(mjUSEDOUBLE)
+  #if defined(__AVX__) && !defined(mjUSESINGLE)
     #define mjUSEAVX
     #include "immintrin.h"
   #endif
@@ -609,7 +609,7 @@ void mju_addToScl(mjtNum* res, const mjtNum* vec, mjtNum scl, int n) {
 void mju_addScl(mjtNum* res, const mjtNum* vec1, const mjtNum* vec2, mjtNum scl, int n) {
   int i = 0;
 
-#if defined(__AVX__) && defined(mjUSEAVX)  && defined(mjUSEDOUBLE)
+#if defined(__AVX__) && defined(mjUSEAVX)  && !defined(mjUSESINGLE)
   int n_4 = n - 4;
 
   // vector part
diff --git a/src/engine/engine_util_blas.h b/src/engine/engine_util_blas.h
index eaac6360..882c9783 100644
--- a/src/engine/engine_util_blas.h
+++ b/src/engine/engine_util_blas.h
@@ -26,7 +26,7 @@ extern "C" {
 
 //------------------------------ standard library functions ----------------------------------------
 
-#ifdef mjUSEDOUBLE
+#if !defined(mjUSESINGLE)
   #define mju_sqrt    sqrt
   #define mju_exp     exp
   #define mju_sin     sin
@@ -59,7 +59,7 @@ extern "C" {
   #define mju_log10   log10f
   #define mju_floor   floorf
   #define mju_ceil    ceilf
-#endif
+#endif  // !defined(mjUSESINGLE)
 
 
 //------------------------------ 3D vector and matrix-vector operations ----------------------------
diff --git a/src/engine/engine_util_sparse_avx.h b/src/engine/engine_util_sparse_avx.h
index f962975c..c7f2f219 100644
--- a/src/engine/engine_util_sparse_avx.h
+++ b/src/engine/engine_util_sparse_avx.h
@@ -16,7 +16,7 @@
 #define MUJOCO_SRC_ENGINE_ENGINE_UTIL_SPARSE_AVX_H_
 
 #ifdef mjUSEPLATFORMSIMD
-#if defined(__AVX__) && defined(mjUSEDOUBLE)
+#if defined(__AVX__) && !defined(mjUSESINGLE)
 
 #define mjUSEAVX
 
@@ -315,7 +315,7 @@ int mju_compare_avx(const int* vec1, const int* vec2, int n) {
   return !memcmp(vec1+i, vec2+i, (n-i)*sizeof(int));
 }
 
-#endif  // defined(__AVX__) && defined(mjUSEDOUBLE)
+#endif  // defined(__AVX__) && !defined(mjUSESINGLE)
 
 #endif  // mjUSEPLATFORMSIMD
 
diff --git a/unity/Runtime/Bindings/MjBindings.cs b/unity/Runtime/Bindings/MjBindings.cs
index 2e791f25..0d551cdd 100644
--- a/unity/Runtime/Bindings/MjBindings.cs
+++ b/unity/Runtime/Bindings/MjBindings.cs
@@ -61,7 +61,6 @@ public const bool THIRD_PARTY_MUJOCO_INCLUDE_MJSPEC_H_ = true;
 public const bool THIRD_PARTY_MUJOCO_INCLUDE_MJTHREAD_H_ = true;
 public const int mjMAXTHREAD = 128;
 public const bool THIRD_PARTY_MUJOCO_INCLUDE_MJTNUM_H_ = true;
-public const bool mjUSEDOUBLE = true;
 public const double mjMINVAL = 1e-15;
 public const bool THIRD_PARTY_MUJOCO_MJUI_H_ = true;
 public const int mjMAXUISECT = 10;