# whisper.cpp v1.9.3

- Product: whisper.cpp (https://whatsnew.fyi/product/whisper-cpp)
- Vendor: ggml
- Date: 2026-08-20
- Version: v1.9.3
- Original notes: https://github.com/ggml-org/whisper.cpp/releases/tag/v1.9.3
- Permalink: https://whatsnew.fyi/product/whisper-cpp/releases/v1.9.3
- Labels: Pre-release

What's New is an index, not a publisher: every entry below links to the vendor's own release notes, which are the authoritative source. Entries are labelled where they are hand-curated sample data, pre-releases, or drawn from a secondary source such as a developer blog.

Reuse: the summaries, labels and curation here are © What's New. Quote freely with attribution and a link back; wholesale republication of the corpus is not permitted — terms: https://whatsnew.fyi/terms. The vendors' own release notes remain their publishers'.

---

- **fixed** — heap out-of-bounds read in log_mel_spectrogram on very short audio
- **fixed** — reject invalid n_dims in tensor header to prevent stack-buffer-overflow on malformed model files
- **changed** — sync ggml to version 0.20.2
- **changed** — sync llama.cpp changes for CUDA, SYCL, Metal, OpenCL, and Vulkan optimizations

> [!NOTE]
> Semantic versioning is still work in progress.
> More info can be found in https://github.com/ggml-org/ggml/discussions/1579

**Nightly build:** [b4938](https://github.com/ggml-org/whisper.cpp/releases/tag/b4938)

##### Change log since v1.9.2

371b5a75 release : v1.9.3 (#4000)
81c19056 cmake : update semver and release process [no ci] (#3996)
4834a232 talk-llama : sync llama.cpp
6b014cfb sync : ggml
8189458d ggml : bump version to 0.20.2 (ggml/1589)
51319a2c CUDA: MMVQ nwarps=8 for bs=1 for dense models on DGX Spark (llama/26843)
71759b7c cuda : skip UMA override for HIP builds (llama/27083)
964bb1b8 ggml : bump version to 0.20.1 (ggml/1587)
4257f470 sycl: fix thread/block count in quantized cpy kernel launches (llama/27160)
0f386138 support OP OPT_STEP_ADAMW, OPT_STEP_SGD (llama/25268)
9a0d1902 vulkan: add SHMEM_STRIDE_PAD/APPLY_SLM_A_RESHAPE for coopmat1 on Intel Xe (llama/25380)
ab71410c fixed indent
9b98e57d Fixed gating logic for problematic Intel driver version
1fe009ca talk-llama : fix build (#0)
733f2819 sync : ggml
667da04a ggml : bump version to 0.20.0 (ggml/1584)
b7ea8b19 ggml : recurrent state rollback for ggml_ssm_scan (llama/26623)
2aef2a00 sycl: fuse mul_mat(gate) + mul_mat(up) + GLU for q4_K dense FFN (llama/26779)
5bc0c852 ggml: force single thread on wasi (llama/25686)
2d05b6ec sycl: fuse the gated-delta-net state writeback cpy (llama/26643)
43cbe415 OpenVINO: Qwen3.5, memory optimization, and test-recurrent-state-rollback (llama/26952)
94d2d35a Support host pinned mem to improve SYCL Host-to-Device Memory Access (llama/26789)
34009e85 metal : add TQ2_0 support (llama/26980)
62031fed ggml-cpu/ops: vectorize flash-attention V-cache F16 to F32 conversion (llama/26947)
ac9a74f1 sycl: remove separate fp32 type promotion in gemm non-oneDNN path (llama/26372)
3425d130 sycl: fuse UNARY(silu|sigmoid|softplus) + MUL (llama/26411)
90d4ed18 sycl : Add DMMV ESIMD Q3_K kernel (llama/26251)
f7011606 sycl : enhance concat to support Q4_0, Q4_1, Q5_0, Q5_1, Q8_0 (llama/26800)
c7be8b33 ggml-hip : remove -funsafe-math-optimizations (llama/26696)
406b1168 ggml : fix arm builds, unused var (llama/26991)
b0e32971 gguf : harden loader against malformed tensor dims and metadata types (llama/25596)
92423afd kleidiai: Add runtime feature detection mechanism for aarch64/kleidiai (llama/26076)
1b93067b opencl: default FA c8 cluster width to 16 on X1E (llama/26433)
63290400 vulkan: add TQ2_0 (ternary) support (llama/25850)
db3687b2 opencl: use flat mv q5_k when weight exceeds image1d_buffer_t limit (llama/26880)
cdbe4551 CUDA: only disable CUDA graphs when mul_mat_id actually needs a stream sync (llama/26802)
030656f2 cuda : add warp-per-row wkv7 kernel for single-token decode (llama/26111)
d9589688 llama: add default load-mode auto, which avoids mmap on iGPUs (llama/26081)
0f60f8e6 ggml-webgpu: fix CI errors from #25025 and #25262 (llama/26566)
a6e2630f opencl: transpose the K tile in local memory for FA prefill kernels (llama/26428)
830ec225 ggml-cpu : fix CPU affinity mask being ignored on Android (llama/26838)
b3bc9046 ggml : require contiguous src for ROLL on CUDA and Metal (llama/25928)
f09a97cf ggml-webgpu : refactor several wgsl files and simplify flash_attn wgsl. (llama/26134)
877761c5 ggml-cpu : fix missing Q5_0 dispatch in SpaceMiT backend (llama/26792)
10791af6 CUDA: fuse rms_norm + mul + rope (+ view + set_rows) (llama/26767)
068d3b0b CUDA: fix thread/block count in quantized cpy kernel launches (llama/26731)
eb3296f2 sycl: coalesce the ssm_conv window loads (llama/26612)
8a5ba01c metal : fix NORM/RMS_NORM for row lengths that leave a partial simdgroup (llama/26708)
06248ca1 cmake : add config version support (ggml/1582)
592feef0 talk-llama : sync llama.cpp
87704926 sync : ggml
84cdcad3 ggml : bump version to 0.19.0 (ggml/1581)
8587ad3b ggml : add aarch64 HWCAP fallbacks and fix fp16 variant detection (llama/25554)
56cb1547 sycl: fix UE4M3 parsing (ll

_[Truncated at 4000 characters — full notes: https://github.com/ggml-org/whisper.cpp/releases/tag/v1.9.3]_