Table 25.3.4 Presents itself like a key update within the branch Table 25.3Designed to fine-tune stability, fix significant bugs, and polish the work introduced in previous versions, this update doesn't bring groundbreaking new features, but rather a set of changes that, in practice, can make a real difference for those who play games on Linux, use Vulkan or OpenGL extensively, or rely on open-source drivers for their daily tasks.
This installment builds upon the foundation of Mesa 25.3, a series that had already introduced A large number of new extensions for Vulkan and OpenGLMesa 25.3.4 includes improvements to drivers such as RADV, ANV, NVK, Zink, Panfrost, V3D, PVR, and many others, as well as advancements in Rusticl, support for Arm Ethos NPUs, and advanced debugging tools. It arrives precisely to bolster this ecosystem: correcting regressions that affected specific games, resolving driver crashes, and fixing compatibility issues that could ruin the experience in both commercial titles and professional applications.
General updates for Table 25.3.4
Initially, Mesa 25.3.4 officially presents itself as a Maintenance version focused on bug fixesIt's the next stop after the 25.3.3 And, as explained in the announcement, its function is to address the errors detected since the previous revision. Although the core of the project continues to offer the implementation of the most important graphics APIs, this version focuses on ensuring that everything already incorporated works correctly.
Regarding supported APIs, the library continues to offer a Full implementation of OpenGL 4.6 and Vulkan 1.4However, the specific level that each user will see depends on the driver they use: the value returned by glGetString(GL_VERSION)By glGetIntegerv(GL_MAJOR_VERSION) y glGetIntegerv(GL_MINOR_VERSION) or through the countryside apiVersion en VkPhysicalDeviceProperties This will vary depending on the actual capabilities of each driver. Some GPUs do not implement all the mandatory features of OpenGL 4.6, so the version advertised in the context may be lower.
An important detail is that OpenGL 4.6 is only exposed when explicitly requested in context creation. In other cases, especially with compatibility contexts, the driver may advertise a lower version based on what it stably supports. Similarly, in Vulkan 1.4, the version number visible to the user adjusts to the level that the driver effectively declares for each physical device.
Expanding hardware support
In terms of hardware, Mesa 25.3 and its sub-versions, including Mesa 25.3.4, continue the trend of Expand the range of supported devices and improve integration with existing platformsThe PVR driver has achieved compatibility with Vulkan 1.2, which is crucial in certain SoCs where this driver is the foundation of the entire graphics stack.
On the Intel side, the ANV driver expands its reach with support for platforms like Wildcat LakeThis ensures that relatively recent systems maintain proper interoperability with Vulkan and continue to benefit from ecosystem updates. In parallel, Intel's integrated graphics receive performance tweaks and fixes for games running through Proton, reducing artifacts and rendering errors.
AMD is not far behind: RADV, its open-source Vulkan driver, continues to receive Ray tracing optimizations and video improvementsIn version 25.3, improvements have been made to video acceleration for the company's GPUs, in addition to integrating adjustments that enhance performance in demanding games and in scenarios where latency is a key factor.
In the NVIDIA open-source ecosystem, the NVK driver has made significant strides, achieving Vulkan 1.4 compatibility on Blackwell GPUs. This positions the open stack more competitively against the proprietary driver, especially for those who want to use only free software without sacrificing modern API capabilities.
Mesa 25.3.4 introduces a new Gallium driver for NPUs and other advanced tools
One of the most striking technical innovations of the 25.3 series is the appearance of a Gallium driver specifically for Arm Ethos NPUsDesigned to work in conjunction with the Teflon framework, this move, while it may sound niche, points toward a future where specialized hardware (such as AI accelerators) is more naturally integrated into GNU/Linux systems.
Another advanced feature that has been added is the possibility of replacing SPIR-V shaders in Vulkan driversThis feature is geared towards debugging and diagnostics: it allows you to swap certain shaders at runtime to better understand what's failing under specific workloads. For engine developers, profiling tools, and game porting studios, this is a significant help.
Symbolic changes and technical cleaning
Among the structural changes of family 25.3, the following stand out: definitive withdrawal of the VDPAU frontendThis move, which had long been considered legacy technology, simplifies the codebase, reduces maintenance, and pushes users toward more modern interfaces for video acceleration, in line with the evolution of the rest of the ecosystem.
The series also incorporates improved support for Vulkan WSI with atomic mode-settingThis provides more precise management of display modes and synchronization. These details are crucial for environments where desktop experience, gaming performance, and consistent frame rates are paramount, such as Steam Deck, gaming PCs, or graphics workstations.
To complete this modernization process, extensions such as VK_KHR_pipeline_binary to ANV and NVKThese features help reduce shader loading times and provide greater stability when managing precompiled pipelines. This is good news for those experiencing stuttering in certain games, as it's part of an effort to smooth out background shader compilation.
Error corrections and specific regressions in Table 25.3.4
Beyond the functional foundation of the 25.3 series, version 25.3.4 introduces a significant list of corrections aimed at very specific problems that had been detectedOne of the most visible issues affects the RADV controller: there was a regression that caused Resident Evil 4 to crash when using QA instruction checks in vkd3d-proton, something especially annoying for those who play this title using Proton.
A solution has also been found Driver crash on an RX Vega 64 when handling a large number of vertex shadersThis case was detected with the OpenGOAL project running Jak and Daxter 1. This type of error usually manifests as complete GPU or graphics system crashes, so its correction has a direct impact on stability.
Another issue addressed affects RADV in the game Kingdom Come: Deliverance 1 on RDNA4 hardware, where the RGP captures lacked cache counters for certain dispatch operationsAdditionally, a situation has been corrected where the NIR lowering process in RADV could cause a segfault if it encountered a radiology query that attempted to proceed before being properly initialized.
Version 25.3.4 also addresses a Regression in the Vulkan driver for Intel integrated GPUsThis further improves reliability in environments where integrated graphics are the primary resource. Additionally, the use of SIMD32 shaders with ray queries has been enabled in the context of ANV/Intel-brw, completing the previous work to enable this type of shader with ray queries.
In the area of memory, a GTT memory leak when running OpenGL games or software on an AMD RX 6600 XTThis could degrade performance in the long term or cause unexpected behavior after prolonged gaming sessions.
In the headless presentation section, the component wsi_common_headless now he makes sure Properly fill in the VkSurfacePresentModeCompatibilityKHR structure when using VK_EXT_headless_surfacePreviously, this information was not completed, which could cause compatibility problems in applications that depended on that data.
A visual glitch has also been fixed in WITCH ON THE HOLY NIGHTWhere some interface elements displayed flickering frames on GPUs with gfx1150/1151 identifiers using RADV. In video, an issue where VA no longer converted correctly from YUV to RGB has been fixed, and a VCN page fault and ring timeout during VAAPI HEVC encoding have been resolved. scale_vaapi on an RX 9060 XT / gfx1200.
In the NIR field, a case is resolved undefined behavior when using a relocated range_minimum_query_tableand a regression in cooperative matrices with RADV is addressed. Furthermore, the compilation option is clarified. gallium-rusticl-enable-drivers to avoid confusion when enabling drivers in Rusticl.
Internal adjustments and developer contributions
The Mesa 25.3.4 changelog reflects the work of a large number of developers, each contributing Specific fixes, stability improvements, and minor internal tweaksAmong the changes signed by Daniel Schürmann, for example, are fixes in aco/lower_to_hw to correct the SGPR operand register classes in subword copies and in 2×16 packing operations, as well as a limitation to prevent the use of two SGPR operands in a single VOP3 instruction before GFX10 in do_pack_2x16().
Dave Airlie adds changes such as a layer option to disable device selection logic In device-select, its use in Zink to disable that selection when necessary, a fix in radv/coopmat related to reference strides and a gallivm setting to swap 1D array coordinates before casting.
Eric Engestrom, for his part, introduces a correction to the option of "meson" (meeting place). “enable-drivers” for RusticlThis avoids inconsistent configurations and limits the integration of RenderDoc in Vulkan runtime and Zink only to platforms where it is officially supported, reducing errors in other environments.
In the area of Mali GPU drivers, Faith Ekstrand makes adjustments such as using bi_emit_collect_to() for load_const in pan/bi, the elimination of calls to blend_emit_descs() when there is no fragment shader in panvk/csf and full flat variable marking in pan/bi, which helps to avoid interpolation errors.
Ian Romanick answers several questions related to NIR algebraic transformations and program management in MesaAmong its changes, the following stand out: the detection of F-strings that were missing the "f" in the algebraic rules, and a correction of segfaults in _mesa_delete_program y _mesa_reference_program_, and the addition of mechanisms to detect precisely those malformed F-strings.
Icenowy Zheng brings a series of fixes focused on Vulkan WSI and the handling of GLES1/2 contexts: destroying images never created in the Vulkan headless backend is avoided, and a workaround for GL_INVALID_OPERATION in GLES 2.0 drawsA GL_INVALID_OPERATION associated with the joint use of GLES1/2 and Kopper is fixed, and another GL_INVALID_OPERATION instance is corrected when releasing a buffer in GLES1/2 contexts.
Lionel Landwerlin's work focuses largely on ray queries and SIMD32 in the Intel stack: adds pending shader release in device_memory_report, ensures that two RT_DISPATCH_GLOBALS structures can be loaded, incorporates divergence handling for ray_query_global_intel In NIR, it prepares the basis for SIMD32 ray queries, allows ray query spilling in SIMD32, handles the lowering of certain additional opcodes, enables topology opcodes in SIMD32, activates SIMD32 compute shaders with ray queries, and corrects the calculation of derivatives in non-32-bit floats.
From AMD's side, Marek Olšák corrects Several problems in radeonsi and NIR utilities: fixes color interpolation when finalize_nir It calls twice, it avoids a segfault in ac/lower_ngg_mesh by accessing outside the limits of out_variables, corrects slightly incorrect assertions in si_shader_ps, fixes an incorrect PS shader key related to sample shading and repairs assertion failures in nir/clip_cull_distance_utils when using GL_EXT_mesh_shader.
Mike Blumenkrantz continues to polish Zink with changes like The correct use of the GENERAL layout in dynamic texture erasing, the elimination of non-operational checks when rewriting erase operations, the creation of new transient images if the number of samples does not match, and the explicit establishment of pipe_resource::next to null when temporary resources are generated.
Samuel Pitoiset provides a series of corrections related to performance counters, SPM, and copy operations On AMD hardware: fixes GRBM broadcasting for global blocks in ac/spm, disable ENABLE_PING_PONG_BIN_ORDER in GFX11.5 within ac/cmdbuf, fixes stencil copies only in GFX9 under ac/sdma, corrects the capture of performance counters using SPM in RADV, delays the reservation of VMID in SQTT at the time of capture, makes RADV meta use 2D array textures for color resolutions with compute and corrects the calculation of cubemap derivatives when the principal axis is negative.
Regarding the WSI layer, Yiwei Zhang is making improvements aimed at dmabuf handling and presentation synchronizationExport routes requiring actual dmabuf support in Zink are hardened, explicit respect for VK_SUBOPTIMAL_KHR returned by WSI image acquisition on Venus is added, a workaround that takes ALIAS into account in the image memory requirements cache is incorporated, the memory requirements cache for alias images on Venus is corrected, the use of the host stage in certain blits to external queues is avoided, and the support conditions for presentation wait and creation of presentation IDs in the Vulkan WSI layer are fixed.
Finally, the user llyyr contributes key improvements for headless environments in Vulkan: ensures that VkSurfacePresentModeCompatibilityKHR is populated correctly, adds a stub for VkSurfacePresentScalingCapabilitiesKHR in non-display configurations, and deploys vkReleaseSwapchainImagesKHR for the headless backend, thus completing the support for swapchains in scenarios without direct video output.
Practical impact on gaming and everyday use
All these changes in Mesa 25.3.4, although in many cases they may seem technical and very specific, translate into a more stable and polished experience for those using open-source driversTitles such as Resident Evil 4, WITCH ON THE HOLY NIGHT, Kingdom Come Deliverance 1 and others that make intensive use of Vulkan, ray tracing, vkd3d-proton or advanced GPU features have seen crashes, graphical artifacts or strange behavior corrected.
For users of Intel, AMD, or NVIDIA GPUs under the open stack, the sum of fixes in ANV, RADV, NVK, Zink, Rusticl, and WSI headless forms an ecosystem where modern games, rendering applications, and general-purpose computing software They run with fewer hiccups. Along with the optimizations introduced during the Mesa 25.3 cycle, there is a noticeable improvement in shader load times, reduced stuttering, and greater latency consistency.
ARM-based platforms, SBCs, and dedicated SoCs also benefit from expanded support in V3D, Panfrost, and PVR, as well as the introduction of the Gallium driver for Arm Ethos NPUs, which opens the door to future scenarios where artificial intelligence and graphics tasks coexist in a more integrated way on LinuxThose working with embedded development or compact devices will see a more solid foundation on which to build.
Overall, Mesa 25.3.4 closes several wounds opened after 25.3.3 and consolidates the path of a 25.3 series that was already full of important new features. Without needing big headlines, the project continues to push the Linux graphics stack towards more mature territory.where every small fix counts so that recent games, modern hardware and advanced tools work more and more reliably.