The daily firehose — everything the toolchain shipped today, already filtered.
// HOW THIS ISSUE IS MADE
We read every release from the 174 tools on our watchlist at the source — GitHub and GitLab release notes, vendor release pages and changelogs, project blogs and feeds, vendor press releases, and the source code behind the tag. Bug-fix-only releases and non-product newsroom noise are dropped; what's left is summarized down to the new capability, how to try it, and any screenshots or videos the release itself published. Every entry links to the sources it was built from.
ONNX Runtime v1.28.0 adds CudaQuantizer, OrtModelPackageApi experimental C API, Linux NPU sysfs discovery, and CUDA EP footprint reductions.
└──▷ GET THIS VERSION
$ git clone --branch v1.28.0 https://github.com/microsoft/onnxruntime.git
# already have the repo? check out this version:$ git checkout v1.28.0
›Adds OrtErrorCode::ORT_DEVICE_RESET to the OrtErrorCode enum, with documentation single-sourced so StatusCode stays in sync.
›Adds model_external_initializers_file_folder_path session option, now honored for file-path model loads.
›Adds CudaQuantizer to onnxruntime.quantization for GPU-side quantization workflows.
›Adds a Python API for HOST_ACCESSIBLEOrtValue allocation.
›Introduces experimental OrtModelPackageApi in the C/C++ experimental API surface, including authoring tools, schema versioning, and folding external_data into session options — note this API may change in future releases.
+14 moreshow less
›CUDA EP: cuDNN and cuFFT are now optional at runtime and nvrtc is no longer linked, significantly reducing the required CUDA redistributable footprint.
›CUDA plugin EP gains support for user compute stream with CUDA graph, kernel sync stream exposed for scratch allocation, and Windows ARM64 packages.
›Adds Linux NPU discovery through sysfs accel devices.
›Adds an API to select the best compiled-model compatibility info from candidate strings.
›Adds crypto support: applications can supply I/O callbacks to an EP, with callback and fallback helpers.
›Implements name-based partitioning in the plugin EP framework.
›Registers Flatten as a Direct8Bit op in the Python QDQ static quantizer.
›Adds Float16, BFloat16, and Float8 support in the TensorArray custom op.
›Adds memory statistics to profiling output.
›Adds EP version logging on inference failure, in the EpDeviceUsage event, and ORT version logging.
›User-supplied external initializers are now used in place when already on the planned device, avoiding unnecessary copies.
›Upgrades to ONNX 1.22.0 and protobuf 6.33.5.
›NPM packages are now published from the CUDA 13 pipeline.
›WebGPU plugin EP version bumped to 0.3.0.
└──▷ BREAKING ON UPGRADE
!CUDA_QUANT_PREPROCESS is now off by default; workflows relying on the previous default-on behavior must explicitly enable it.
!The TensorRT fused causal attention kernels were removed from the CUDA EP; models depending on them will need an alternative path.
!The dynamic WGSL generator (duktape/Node) path was removed; the Python wgsl-gen implementation is now the only supported path.
!OrtModelPackageApi now lives in the experimental C API and may change in future releases; callers should treat its signatures as unstable.
!SkipLayerNorm strict mode is deprecated and may be removed in a future release.
Lightweight coding agent that runs in your terminal
Codex CLI alpha adds configurable code-mode host fallback, hardened network approval concurrency, and richer Responses Lite metadata.
└──▷ GET THIS VERSION
$ git clone --branch rust-v0.146.0-alpha.10 https://github.com/openai/codex.git
# already have the repo? check out this version:$ git checkout rust-v0.146.0-alpha.10
›Adds disable_in_process_fallback key inside a features.code_mode_host configuration table, letting operators suppress the embedded-V8 fallback so that a failed standalone host surfaces as tool output instead of silently falling back.
›Preserves backward compatibility: features.code_mode_host still accepts a plain boolean toggle in addition to the new table form.
›Adds code_mode_tool_names field to Responses Lite turn metadata, mapping each normalized code-mode identifier to its structured ToolName (omitted from external MCP server metadata and non-Lite requests).
›Scopes pending network approvals to a turn and execution, coalesces duplicate requests within one execution, and cancels in-flight Guardian reviews when their owner is dropped — eliminating races between concurrent allow/deny decisions.
PydanticAI v2.18.0 adds AdvisorTool for Anthropic/OpenRouter, BedrockMantleProvider, multi-region Google Cloud, and external web access for OpenAI WebSearchTool.
└──▷ GET THIS VERSION
$ git clone --branch v2.18.0 https://github.com/pydantic/pydantic-ai.git
# already have the repo? check out this version:$ git checkout v2.18.0
└──▷ USE IT
Route Bedrock requests through the Mantle provider for managed AWS Bedrock access.
python
from pydantic_ai.providers.bedrock_mantle import BedrockMantleProvider
provider = BedrockMantleProvider()
›Adds external_web_access option to WebSearchTool for OpenAI Responses API, enabling built-in web search without a separate tool.
›Adds BedrockMantleProvider for AWS Bedrock Mantle integration, with normalized response-scoped tool-call IDs.
›Extends AdvisorTool support to Anthropic and OpenRouter providers.
›Adds 'us' and 'eu' multi-region location values to GoogleCloudProvider location type.
›Adds --speculative-algorithm DSPARK flag for DSpark confidence-driven speculative decoding, with --speculative-dspark-block-size to tune block size; requires SGLANG_RAGGED_VERIFY_MODE=compact environment variable — reaches 383.7 tok/s at accept length ~5 on DeepSeek-V4-Pro TP8.
›Adds --enable-dsa-cache-layer-split flag (requires --enable-prefill-cp --cp-strategy interleave) for GLM-5.2 DSA cache layer sharding across CP ranks, cutting per-rank KV memory ~74% (0.77 to 0.20 GB/rank) at 8192 tokens.
›Adds --enable-gdn-replayssm-spec flag for ReplaySSM Ring Spec-Verify (GDN), reducing speculative scratch from 11.5 GB to 1.8 GB per GPU (6.4x smaller) on Qwen3.5-35B-A3B at TP1; tune ring size with --linear-replayssm-cache-len. Requires --speculative-eagle-topk in {None, 1}.
›Adds native support for Inkling, a 975B-parameter multimodal MoE with 1M-token context, NVFP4 MoE, optional vision/audio towers, and native MTP — verified on Blackwell TP4/TP8, H200, and AMD MI350X/MI355X.
›Makes UnifiedRadixTree the default cache tree for SWA, Mamba, and DSA models, with Replay SSM and Mamba int8 checkpoints synced onto it and cache-hit state reset scoped to only the used state.
+6 moreshow less
›Adds support for new models: LongCat 2.0 FP8, JetBrains Mellum v2, Pi0.5 (VLA), and LongLive 2.0 (diffusion).
›Adds linear attention decode kernel (recurrent_kda) for Blackwell SM100, running at 29.6 µs vs 36.8 µs for Triton at B=64, with 1.35x throughput at B=256.
›Adds DeepSeek-V4 Wint4Abf16 and Wint4Afp8 quantization support.
›Integrates Q8KV8 FP8 Sparse MLA Prefill into the DSA backend for DeepSeek-V3.2.
›Fuses state I/O into the chunk-h kernel for GDN/KDA CuteDSL prefill.
└──▷ BREAKING ON UPGRADE
!The experimental QServe (QoQ) W4A8 and FBGEMM FP8 quantization paths are removed; any server launch using these will fail.
!CUTLASS FP8 blockwise kernels are removed for SM90/SM100; SM120 is moved to JIT.
!--fp4-gemm-backend cutlass is removed along with in-tree NVFP4 JIT kernels; NVFP4 GEMM now requires FlashInfer. Use auto, which picks flashinfer_cutedsl on SM100 and flashinfer_cutlass on SM120.
!UnifiedRadixTree is now the default for SWA, Mamba, and DSA models, changing cache behavior on those architectures.
!Chunked input-logprob processing is now on by default to cap peak memory.
!FA3 sparse mask kernels are off by default.
!Kernels are relocated to the sglang.kernels namespace; code reaching past public wrappers to internal import paths must update.
!num_tokens_per_bs is renamed to num_tokens_per_req across spec-decoding runners.
!--enable-deepep-waterfill is renamed to --enable-waterfill with no deprecated alias; existing launch commands will fail with unrecognized arguments.
!--optimistic-prefill-retries is renamed to --optimistic-prefill-attempts with no deprecated alias.
!The SGLang-Diffusion post-training rollout endpoint now returns application/msgpack instead of JSON, with tensors as raw msgpack bytes; tensor_to_base64 / base64_to_tensor are replaced by tensor_to_bytes / bytes_to_tensor — RL rollout consumers must be upgraded in lockstep with the server.
THE AI TOOLCHAIN