For the complete documentation index, see llms.txt. Markdown versions of all pages are available by appending .md to any URL (e.g. /get-started.md).
MAX nightly
This version is still a work in progress.
Highlights
Documentation
Models
-
Qwen 3.5 models served with MTP or DFlash2 speculative decoding now read the images and videos in a request. They used to answer without them.
-
GLM-5.1, GLM-5.2, and GLM-5.3 now default
clear_thinkingtoFalse, which keeps thereasoning_contenta client sends back for earlier turns in the prompt. GLM-5.3 previously defaulted it toTrue, and GLM-5.1's and GLM-5.2's templates clear earlier reasoning when it is unset. This matches Z.ai's Coding Plan endpoint, and Z.ai's thinking-mode guide recommends it for coding and agent workloads. Passclear_thinking=Truein the chat template options to drop earlier reasoning. -
FLUX.2 models now generate more than one image per request, so
num_imagesabove 1 inmax serveandnum_images_per_promptingenerate_images()andAutoModel.generate()work for them. Such a request used to fail in the first denoising step. -
InternVL and Gemma 3 multimodal models now return log probabilities, so
logprobsinmax serveandoutput_logprobsingenerate()work for them. -
Added support for
poolside/Laguna-S-2.1-NVFP4on theLagunaForCausalLMarchitecture. The checkpoint interleaves 12 full-attention layers with 36 sliding-window layers on a 512-token window, gates attention output per head rather than per element, and mixes expert precisions: NVFP4 routed experts in layers 1-39 and bfloat16 ones in layers 40-47, with the dense first layer and every shared expert in bfloat16 throughout. Attention type, query-head count, RoPE scaling and MLP precision are now all resolved per layer, sopoolside/Laguna-M.1-NVFP4keeps its uniform layout. -
Added support for
LiquidAI/LFM3-1.7BandLiquidAI/LFM3-1.7B-Baseon the newLfm3ForCausalLMarchitecture. -
Added support for
LiquidAI/LFM3-VL-2Bon theLfm3VlForConditionalGenerationarchitecture, with image and video input. -
Added support for
LiquidAI/LFM3-Omni-2Bon theLfm3OmniForConditionalGenerationarchitecture, with text, image and video input.
Inference server
-
max servenow sends about ten tokens per streamed chunk instead of one, which cuts network traffic and per-chunk overhead without changing time to first token.MAX_SERVE_STREAM_MIN_CHUNK_TOKENS=1restores the old behavior. -
max serve,max generate,AutoModel, andAutoPipelinenow prefer ModuleV3 when an architecture has both ModuleV3 and graph API implementations. Pass--no-prefer-module-v3on the command line orprefer_module_v3=Falseto use the graph API implementation. -
Serving metric names now use
max.*instead ofmaxserve.*. Batch instruments usemax.batch.*, such asmax.batch.size. Prometheus exposes the correspondingmax_*families. Update collector allowlists, dashboards, and alerts with the deployment; existing metric history stays under its previous names. Units, labels, histogram boundaries, and benchmark result fields are unchanged. -
max servenow rejects an image generation request that asks for more than 4 images with a 422 error namingnum_images, the same response as other out-of-range image options.generate_images()andAutoModel.generate()raise aValueErrorfornum_images_per_promptabove 4 before any prompt runs. A request had no upper limit before. -
/v1/completionsnow returns a 400 error forechowithlogprobson a server started without--enable-echo, when the model computes log probabilities. The request used to crash the model worker. Other models still return no log probabilities, andechowithoutlogprobshas no effect. -
Qwen 3.5 now passes every
chat_template_kwargsentry to its chat template, such asadd_generation_promptorpreserve_thinking. It used to keep onlyenable_thinking, which still defaults to on. -
MAX Serve's span attributes now sit under the entity they describe:
max.request_idis nowmax.request.id,max.batch_idismax.batch.id, and the other request and batch attributes move undermax.request.*andmax.batch.*.kernel_agg:<name>spans carry the kernel name inmax.kernel.opinstead ofmax.kernel.name; per-kernel spans keepmax.kernel.name. Update trace queries that filter on the old names.
Server metrics
-
maxserve_batch_execution_time_millisecondsnow measures the executor call when overlap is off. The former full scheduler-step duration is available asmaxserve_scheduler_batch_step_time_milliseconds. Batch throughput uses the executor interval and counts returned output tokens, including prefill outputs and accepted speculative tokens. See the batch timing contract for overlap timing. Intermediate prefill chunks returning no tokens now contribute zero to the CE generation-throughput histogram, shifting its distribution on chunked-prefill workloads. The benchmark harness'sprefill_batch_execution_time_msanddecode_batch_execution_time_mscolumns usemaxserve_batch_execution_time_milliseconds, so MAX's synchronous values now exclude scheduler work; comparisons with older builds are not like for like. -
Added
maxserve_request_client_disconnect_cancellations_total, which counts streamed responses cut short by a client disconnect and completion requests abandoned while waiting for a result, split byphase. A mid-stream disconnect was previously invisible:maxserve_request_count_totalrecords it under the status code already sent.
max CLI
-
PipelineArgssub-config flags are now flat, matchingmax serve's spellings everywhere the class is bound with cyclopts. The benchmark engine's--pipeline.runtime.max-batch-sizebecomes--pipeline.max-batch-size,--pipeline.kv-cache.enable-prefix-cachingbecomes--pipeline.enable-prefix-caching, and--pipeline.profiling.gpu-profilingbecomes--pipeline.gpu-profiling; theMODULAR_PIPELINE_*environment variable names change with them (MODULAR_PIPELINE_MAX_BATCH_SIZE, notMODULAR_PIPELINE_RUNTIME_MAX_BATCH_SIZE). The experimental cascade serve CLI (//mach:cascade) now accepts the fullmax serveflag surface, including--config-filerecipes and the--model,--devices, and--use-dummy-weightsaliases. -
max generateruns image generation models such as FLUX.2 and saves the images to--output, which defaults toimage.pngand whose suffix picks the format. A path without a suffix is a directory, created with its parents if missing. It adds--negative-prompt,--height,--width,--num-inference-steps,--guidance-scale,--true-cfg-scale,--strengthand--num-images-per-promptfor them, and takes the input image from--image_url. A flag the user set for the other kind of model, including--max-batch-sizeand--device-graph-captureon an image model, raises an error before anything loads. Text models run as before. -
max benchmarkrenames--warmup-to-steady-stateto--warmup-multiturn, since it applies only to multi-turn workloads and is unrelated to the steady-state measurement window. The old flag, its workload and config file keys, andMODULAR_WARMUP_TO_STEADY_STATEstill work and log a deprecation warning. Result files record the option aswarmup_multiturnunderclient_args.
Python API
-
AutoModel(eager=True),AutoPipeline(eager=True)andmax generate --eagerrun the decoder op by op instead of as a compiled graph, sopdbstopped inside a module'sforward()sees real values on every step. It is for debugging and much slower than the compiled graph. It runs ModuleV3 text generation architectures that register aDECODERmodule, on one device or one device per data-parallel replica. See the pdb guide. -
The eager
Moduleframework moved frommax.experimental.nntomax.nn, and the graph-based layers thatmax.nnheld moved tomax.nn.legacy.max.nn.legacyis deprecated and will be removed in MAX 27.1. Code that uses it gets aDeprecationWarningnaming the line that used it. To keep using the graph-based layers until then, changefrom max.nn import ...tofrom max.nn.legacy import .... -
A custom op whose Mojo kernel is registered wrong, or doesn't match its call, now raises
max.graph.KernelRegistrationError, aValueErrorsubclass. The message points at the Mojo code the way the Mojo compiler does: the file, line and column, the source line and a caret. A custom op package's kernels are now checked when it loads, so a misplacedDeviceContextargument fails there instead of while compiling the generated code. -
Source tracebacks (
InferenceSession.debug.source_tracebacks) now stop at your code's call into the graph API, so errors point at the lines that created each op rather than at MAX's implementation of it. Setsource_tracebacks_hide_internal = Falseto keep every frame. Errors thatModel.executeraises are nowmax.engine.ModelExecutionError, aRuntimeErrorsubclass that is also aValueError, so existingexcept ValueErrorhandlers keep working. -
max.nn.legacy.Transformer,DistributedTransformer,logits_postprocess()anddistributed_logits_postprocess()return amax.nn.legacy.graph_groups.DecoderOutputsof named groups instead of a tuple. Read the logits asoutputs.logits["logits"]rather thanoutputs[1], and pass the outputs toGraph.output(outputs)rather than unpacking them. The graph's outputs keep their order, andGraph.output_treereturns them by name. -
ops.matmul()takes anout_dtypekeyword argument that sets the result dtype. The inputs keep their own dtypes and are no longer cast to a common one, sobfloat16inputs without_dtype=DType.float32keep the float32 accumulator. Withoutout_dtype,ops.matmul(), the@operator andmax.functional'smatmul()require both inputs to share a dtype, which the result keeps, somax.nn.legacy.Linearraises when its input and weight dtypes differ.max.nn.Linearcasts its input to the weight's dtype, so a float32 input to a bfloat16Linearnow gives bfloat16 instead of float32. -
Added
max.driver.Buffer.memset(value)to fill a contiguous buffer or slice in place with a byte value, without allocating a source buffer. -
Added
max.driver.Buffer.memset(value)to fill a contiguous buffer or slice in place with a byte value, without allocating a source buffer. -
Added
ops.allreduce_rms_normandmax.functional.allreduce_rms_norm, a fused allreduce + RMSNorm collective for bfloat16 2-D tensors: every device ends up with the RMS norm of the summed tensor, computed in one launch. The norm runs on the unrounded float32 sum, so it is not bit-identical toallreduce_sumfollowed byrms_norm, which reads the sum rounded back to bfloat16. Themax.functionalform falls back to that two-launch path for inputs that are not 2-D bfloat16 and for meshes without signal buffers. -
Added
ops.allreduce_post_norm_residual_rms_normandmax.functional.allreduce_post_norm_residual_rms_norm, which run a transformer block's branch allreduce, its post-norm, the weighted join to the residual stream and the next pre-norm in one launch per device. They take bfloat16 2-D tensors and need P2P access between the GPUs. On NVIDIA GPUs,allreduce_rms_normnow also fuses small payloads into one launch. -
Added the
ops.CoordinateTransformModeandops.RoundModeenums.ops.resize_linear,ops.resize_nearest, and theirmax.functionalcounterparts take them forcoordinate_transform_modeandround_modein place of the integer codes 0-3. The integer codes still work. -
ops.resize_bicubic()and itsmax.functionalcounterpart take anantialiaskeyword argument. Withantialias=True, the resize uses the filter of Pillow'sBICUBICand matches PyTorch'sinterpolate(mode="bicubic", antialias=True)to float32 rounding. The antialiased resize runs on the CPU, on a float32 input, or on a uint8 input that it resizes to uint8 bit for bit as Pillow'sBICUBICdoes.ops.resize_bicubic()also takes its output size as a runtime int64 tensor, without_dimsdeclaring the output's dimensions. -
Added
max.driver.float32_to_bfloat16_as_uint16(),max.driver.bfloat16_as_uint16_to_float32(), andmax.driver.float32_array_to_buffer()for moving bfloat16 data between NumPy and aBuffer.max.pipelines.libstill re-exportsfloat32_to_bfloat16_as_uint16()andfloat32_array_to_buffer(). -
Removed the
max.pipelines.lib.bfloat16_utilsmodule. Import its functions frommax.driverinstead. -
Removed
GraphPipelineModel,ModuleV3PipelineModelandMultiGraphPipelineModelWithKVCachefrommax.pipelines.lib. SubclassPipelineModel, return the module fromgraph_specs()underGraphRole.ENCODER, and load it withload_graphs()instead. -
Removed the
LLMclass, deprecated in 26.5 whenmax.entrypointsbecame private. To generate in-process, useAutoModelorAutoPipelinefrommax.pipelines. -
Buffer.to(device)for a buffer on another queue of that device requires anafterevent and no longer syncs implicitly. Mojo'sreassign_ownership_to()is removed; useenqueue_wait_for()and keep buffers alive until synced. -
Removed
max.nn.legacy.FusedMLP. Usemax.nn.legacy.MLP, which takes the separategate_projandup_projweights a checkpoint ships and fuses their matmuls in the graph. -
max.nn.legacy.kv_cachemoved tomax.nn.kv_cache. Changefrom max.nn.legacy.kv_cache import ...tofrom max.nn.kv_cache import .... -
Removed
max.nn.legacy.kernels,max.nn.legacy.state_space,max.nn.legacy.data_parallelism,max.nn.legacy.quant_ops,max.nn.legacy.moe.quant_strategy,max.nn.legacy.sampling.penalties, andmax.nn.legacy.comm.ep.ep_kernels. The graph ops they held are now internal to MAX.max.nn.kernelsprovides the kernelsmax.nnmodules use, andLogitPenaltiesstays importable frommax.nn.legacy.sampling. -
Removed
max_single_alloc_size()andrun_healthcheck()from the MojoDeviceContext. On NVIDIA and AMD GPUs the health check still runs when a device context is created. Python'sDevice.max_single_alloc_sizeis unchanged. -
generate(),GenerateIterator,AutoModel.generate()andAutoModel.generate_iter()takemessagesandimages, so they run chat and image prompts on models such as Gemma 3 and InternVL.GenerateResult.prompt_lengthsgives each prompt's length.AutoModelturns the vision encoder cache off by default. -
Added
AutoProcessortomax.pipelines, which turns chat conversations with inline images into themessagesandimagesthatAutoModel.generate()takes. -
AutoPipelinenow takes a chat conversation, or a list of them, with images inline, and returns only the new text for each one. -
generate(),GenerateIterator,AutoModel.generate(),AutoModel.generate_iter()andAutoPipelinetakechat_template_kwargs, which reach the chat template of everymessagesrow, such as{"enable_thinking": False}to turn off Qwen 3.5's thinking. -
imagesongenerate(),GenerateIterator,AutoModel.generate()andAutoModel.generate_iter()takes PIL images as well as encoded bytes, and the tokenizer uses a PIL image without decoding it again.AutoProcessornow passes a copy of each PIL image instead of encoding it. -
AutoProcessorandAutoPipelinenow read images from http(s) URLs,data:URIs and base64 data, including OpenAIimage_urlparts. URLs are fetched with a 30 second timeout and a 100 MiB limit per image. -
Errors from
AutoModel.generate()andAutoModel.generate_iter()now nameAutoModel, notgenerate(). An image part that holds its own image, or animage_urlpart, now raisesValueErrorpointing toAutoProcessor, even whenimagesgives the row the right count. -
Added
generate_images()inmax.pipelines.lib.generate_images, which runs an image generation pipeline such as FLUX.2 in your own process and returns one list of PIL images per prompt, in prompt order. Each prompt runs as the requestmax servebuilds for it, so the model's defaults are the same as in serve. -
AutoModelnow loads image generation models such as FLUX.2, andAutoModel.generate(prompt=...)runs them and returns aGenerateImagesResult. Text models keep the samegenerate()call and result. An argument for the other kind of model raises aValueErrorthat names it. Image models needmax_batch_size=1and no device graph capture.AutoPipelineruns image models too, throughAutoModel.generate(prompt=...), and returns a prompt's list of images. On an image model,AutoModel.generate_iter()points toAutoModel.generate(prompt=...), andgenerate()points togenerate_images(). -
generate_images(),AutoPipelineand the image path ofAutoModel.generate()takeimages, one list of input images per prompt as for textgenerate(), each image encoded bytes or a PIL image. A list holds one image, or none for a text-only prompt. FLUX.2 and Z-Image run image to image this way. They also takestrength, which Z-Image reads and FLUX.2 ignores.ZImagePipelinenow lists image input. -
generate()andGenerateIteratorraise the sameTypeErrorfor a bare image or bytes passed asimagesas for any otherimagesthat is not one list of images per conversation, instead of a Python iteration error. -
Removed the
blocking_sync,disable_timingandinterprocessparameters from the MojoDeviceContext.create_event(), along withEventFlags. Events never record timing; useDeviceContext.execution_time()to time kernels. -
Removed the
per_layer_buffersKV cache option, which gave each layer its own buffer to get past a device's per-allocation size limit, along with thekv_blocks_per_layerandkv_scales_per_layerfields ofKVCacheInputsPerDevice. A KV cache is again one buffer per device, capped at that limit. -
Added
Device.architecture_version, which returns anArchitectureVersionwithmajor,minorandsteppingfields so code can compare architectures instead of parsingarchitecture_name: the compute capability on NVIDIA GPUs (sm_100ais 10.0.0), the processor version on AMD GPUs (gfx950is 9.5.0), and the Apple silicon generation on Apple GPUs (an M5 is 5.0.0). Versions compare for equality only, since a higher version is not always a newer part. A virtual device for compile-only mode whose target is an architecture name, such assm_100aorgfx950, reports the version of that architecture.
C API
Kernels and GPU programming
-
A custom op's
executemethod no longer acceptsManagedTensorSlicearguments (InputTensor,OutputTensorand their fused, mutable and variadic variants). Loading a kernel package that uses them fails with an error at the offending argument. Declare the arguments with tensor traits instead, for exampleSome[Input & Tensor],Some[Output & Tensor], orTensorTuplefor variadic operands. -
DeviceStream.wait_for_host_value()inmax.gpu.hostis deprecated; useDeviceStream.wait_for_flag()instead. -
Built-in custom ops are now named after their class instead of
mo., which is kept for the graph compiler's own ops. For example,mo.launch_host_funcandmo.inplace_memcpyare nowruntime.launch_host_funcandruntime.inplace_memcpy, andmo.matmul_swigluis nowmatmul.swiglu. Update the name passed toops.custom()in graphs that call these ops directly. -
On AMD MI355X (CDNA4), the fused gated activation now quantizes to MXFP4 as well as MXFP8, so a
uint8/float8_e8m0fnurequest fromswiglu_quantize_dynamic_block_scaledruns as one kernel. Such a request was rejected on AMD before; quantizing MXFP4 without the fusion writes the bfloat16[M, N]activation of the[gate | up]input out with a SwiGLU kernel and reads it back in an ungated quantize, and the fused kernel never materializes it. -
Enabled automatic fusion of eligible bfloat16 matrix multiplications and SwiGLU activations on NVIDIA B200 GPUs.
-
The deprecated
LayoutTensorandLayoutTensorItertypes have been removed from thelayoutpackage, along with theto_layout_tensor()andlt_to_tt()conversions. UseTileTensorandTileTensorIterinstead.ThreadScopenow lives inlayout.tile_io. -
The legacy
LayoutandIntTupletypes have been removed from thelayoutpackage, along with the originallayout.layoutandlayout.int_tuplemodules, theto_layout(),LTToTTLayoutandprint_layout()bridges, and the legacytile_layout_k_major(),tile_layout_mn_major()andtile_sf_layout_k_major()constructors. Use the typed layouts in the newlayout.layout(Layout, also exported asMixedLayout) for layout values,Coordfor shapes, strides and indices, and the*_typedtile layout constructors inlayout.tensor_core_async.wgmma_desc_sbo_lbo()andwgmma_canonical_strides()replacetile_to_descriptor(), andprint()accepts a typed layout directly. -
TileTensor.tile()on a nested layout now indexes each mode first leaf first, ascrd2idxand the CuTelocal_tileoperation do: tilecof sizetstarts at elementc * tof the mode, and the tile keeps the leading leaves it spans, with their strides. Before, it treated the first leaf of each mode as the tile grid and the second as the fragment. To extract fragments, put the fragment first in each mode, asblocked_product()does:((3, 2), (4, 2))rather than((2, 3), (2, 4)). A tile size that doesn't split a nested mode on leaf boundaries is now a compile-time error. -
TileTensor.load_linear()andTileTensor.store_linear()have been removed. UseTileTensor.load()andTileTensor.store()with aCoord, which also work on nested layouts:t.load_linear[4](IndexList[2](r, c))becomest.load[4](Coord(r, c)). On CPU,load()andstore()default to the element's alignment rather than the full vector's, so passalignment=to keep a wider one. -
@extensibility.registertakes a platform as a second argument that selects the devices a custom kernel supports, built from thetype,api,arch, andmodelplatforms and theCPU,ACCELERATOR,CUDA,HIP, andMETALhelpers, combined with&and|into a named alias:comptime TCGEN05 = ACCELERATOR & CUDA & (arch("sm_100a") | arch("sm_103a")), then@extensibility.register("my_op", TCGEN05). When several registrations of an op match a device, the graph compiler now selects the one whose set of devices is a strict subset of the sets of all the others, and reports an error if no registration is. Before, it selected the registration that set the most device fields. Two registrations of an op for the same set of devices are an error. Thetype,api,arch, andmodelarguments ofregisterare removed: replacetype="gpu", api="cuda"withCUDA, and in general with the&of the matching platforms. -
Removed the
max,min,sum,productandmeanoverloads inmax.algorithmthat take an input shape, aninput_fn, anoutput_fnand areduce_dim. To reduce along an axis, use the graph ops (ops.sum,ops.max,ops.min,ops.mean,ops.prod), or callmap_reduceonce per row. TheSpanoverloads and the 1-Dsumandmeanoverloads that take a length and aninput_fnare unchanged. -
Added
DeviceContext.device(), which returns aDeviceview of the accelerator with anidfield andget_name(),get_attribute(),is_host_unified(),get_arch()andcan_access_peer()methods.DeviceContext.id(),name(),get_attribute(),is_host_unified()andcan_access()are deprecated: writectx.device().id,ctx.device().get_name(),ctx.device().can_access_peer(peer.device())and so on. -
DeviceContext.enqueue_function()accepts aGridBlockLaunchConfig(grid, block, dynamic shared memory) orClusterLaunchConfig(plus thread block clusters) frommax.gpu.hostas a keyword-onlyexecution_configargument after the kernel arguments. A cluster launch on a GPU without clusters fails to compile; generic code checksconforms_to(DeviceContext[].ExecutionConfigType, ClusterExecutionConfig). On NVIDIA GPUs,GridBlockAttributeLaunchConfigandClusterAttributeLaunchConfigfrommax.gpu.host.nvidiaalso carry launch attributes, such as the onespdl_launch_attributes()returns for programmatic dependent launch. -
A bare
DeviceContextis now an unbound parameterized type, not a concrete type.Code written expecting a concrete
DeviceContextcan change toDeviceContext[]to bind the default parameter values.Prefer to accept generic
DeviceContextas input instead of restricting an implementation to the default parameter values. -
stencil()andstencil_gpu()inmax.algorithmnow take aCoordwhere they took anIndexList[rank]: theshapeandinput_shapearguments, and the point passed to theload_fn,compute_fnandcompute_finalize_fncallbacks. A shapeCoordkeeps its static dimensions on the way to the kernel. Wrap an existingIndexListshape withCoord(shape).map_fnstill takes and returnsIndexList. The point holds the same values as before:# was: def load_fn[...](point: IndexList[rank, ...]) -> ... def load_fn[simd_width: Int, dtype: DType](point: Coord) -> SIMD[ dtype, simd_width ]: return input.load[width=simd_width](point)
Tools
- Added the ability to install
max-serveandmax-benchmarkas Conda extras, usingmax[extras=[serve]]andmax[extras=[benchmark]]respectively.max[extras=[all]]installs both plusmojo.
Fixes
-
Reusing precompiled MEFs (
InferenceSession(precompiled_mefs=...),--precompiled-mefs, andmax.tensor.support.set_precompiled_mefs) now checks that each artifact was built from the graph claiming it: its body, the contents of its kernel libraries, its compile options, and the MAX build that exported it. Before, only the graph's name and its input and output types were compared, so editing a graph or a custom op and reusing an artifact silently ran the old program. A mismatch raisesArtifactBodyMismatch. A caller that knowingly builds different graphs than it recorded, such as a test harness compiling under virtual devices, passesMefStore.for_import(..., verify_body=False),InferenceSession(verify_body= False), or--no-precompiled-mefs-verify-bodyto skip the check. -
Loading a model on a device that reports no free memory now fails with a "Model size exceeds available memory" error. Memory planning used to treat 0 as unmeasured and size against the device's total memory, so a full host could crash during loading instead.
-
max servenow runs prompts given as token ids, with no images, on Gemma 3, InternVL, Idefics3, Pixtral and Qwen 2.5 VL. These requests used to fail in the tokenizer. -
max servenow returns a 400 for a prompt given as token ids, on completions, chatprompt_tokensor embeddings, when an id is below 0 or past the model's vocabulary. Such a request used to fail its whole batch.generate(),GenerateIteratorandAutoModel.generate()raiseInputErrorfor such an id ininput_ids. Embeddings requests the tokenizer rejects, such as a too-long prompt, now return a 400, not a 500. -
The free memory that CPU and Apple GPU devices report now accounts for memory used by other processes. Before, MAX could start a model that didn't fit in memory, which on macOS could exhaust swap and crash the machine.
-
bfloat16models now run on CPU, so--devices=cpuworks with a model whose default encoding isbfloat16without--quantization-encoding=float32. Such a model used to fail at startup with "The encoding 'bfloat16' is not compatible with the selected device type 'cpu'" (#7611).
Mojo language
For all the updates to the Mojo language, standard library, and tools, see the Mojo release notes.