v0.27.0

Major Changes

• Windows arm64 cpu support by @jihadkhawaja in #1226
• Update For Qwen3.5/Gemma4 Support by @martindevans in #1356
• Context Overflow Fix by @zsogitbe in #1373

Bug Fixes

• Fix flaky unit tests by @m0nsky in #1350
• Fix failing workflow caused by new python version by @m0nsky in #1329
• Fix/remove standalone metal shader by @m0nsky in #1331
• Replace -DLLAMA_CURL=OFF with -DLLAMA_OPENSSL=OFF by @m0nsky in #1333
• Linux RUNPATH by @m0nsky in #1330
• Fix compile.yml regressions + switch windows arm64 build to clang by @m0nsky in #1334
• Removed outdated reference to llava_shared.dll for Windows arm64 target by @martindevans in #1339
• Disable shared object versioning by @m0nsky in #1346
• Fix musl detection using RID instead of distro name by @m0nsky in #1347
• bugfix: bugfix: LLamaWeights.NativeHandle.LoadLoraFromFile need read & write access to lora file by @harry330 in #1368

Other Changes

• Expand CI matrix, new platforms + renames by @m0nsky in #1348
• Remove setting hardcoded value for n_seq_max in IContextParamsExtensions by @Ed-Pavlov in #1354
• docs: add FAQ entry for "unknown model architecture" error by @Klivess in #1371

New Contributors

• @jihadkhawaja made their first contribution in #1226
• @Ed-Pavlov made their first contribution in #1354
• @harry330 made their first contribution in #1368
• @Klivess made their first contribution in #1371

Full Changelog: v0.26.0...v0.27.0

v0.26.0

Major Changes

• Add support for gemma 3n by @krisbiradar in #1248
• Mtmd implementation by @SignalRT in #1261

Bug Fixes

• Respect ChatOptions.Instructions in LlamaExecutorChatClient by @stephentoub in #1256
• set RPATH to "@loader_path" / "$ORIGIN" to ensure executables and dynamic libraries search for dependencies in their origin directory by @SignalRT in #1279
• Fixes Extended grammar optimization mode would crash if TopK sampling is set to zero. by @SerialKicked in #1282
• Fix issue #382 (multiple publish output files with same relative path) by @LucaMaccarini in #1234

Other Changes

• Optimization - A queue with fixed storage size backed by a circular buffer by @SignalRT in #1262
• Fix some warnings. by @Lamothe in #1274
• Accept CancellationToken in ChatSession.InitializeSessionFromHistoryAsync by @Cryptoc1 in #1260
• Downgrade compile.yml from macos-latest to macos-14 by @m0nsky in #1302
• Update Vulkan from 1.3.261.1 to 1.4.335.0 and fix download url by @m0nsky in #1303
• Update BinaryReleaseId by @m0nsky in #1304
• Add KaiROS AI as example application by @avikeid2007 in #1313

New Contributors

• @Cryptoc1 made their first contribution in #1260
• @krisbiradar made their first contribution in #1248
• @LucaMaccarini made their first contribution in #1234
• @avikeid2007 made their first contribution in #1313

Full Changelog: v0.25.0...v0.26.0

v0.25.0

Major Changes

• Bianries Update July 2025 by @martindevans in #1225
• docs: update the API references. by @AsakusaRinne in #1213, #1214

Bug Fixes

• Conversation -1 instead of _end by @martindevans in #1197
• Preserve cut off generated text by @OoLunar in #1202
• Fix for getting CUDA Version inside nvidia/cuda containers by @Crelex in #1222

Other Changes

• Update to stable Microsoft.Extensions.AI.Abstractions by @stephentoub in #1190
• Memory efficient context handling by @zsogitbe in #1183
• Expose DecodeSpecialTokens through IInferenceParams for StatelessExecutor by @jacob-mink-1996 in #1203
• Unseal the DefaultSamplingPipeline by @m0nsky in #1208
• Updated main workflow from windows-2019 to windows-2022 by @martindevans in #1216, #1218
• Multi modal (Just libraries) by @SignalRT in #1220

New Contributors

• @omahs made their first contribution in #1196
• @OoLunar made their first contribution in #1202
• @leopardracer made their first contribution in #1206
• @jacob-mink-1996 made their first contribution in #1203
• @kilavvy made their first contribution in #1207
• @vtjl10 made their first contribution in #1209
• @zeevick10 made their first contribution in #1210
• @maximevtush made their first contribution in #1211
• @Crelex made their first contribution in #1222

Full Changelog: v0.24.0...v0.25.0

v0.24.0

Major Changes

• May Binary Update by @martindevans, @AmSmart and @nipeone in #1179 / #1163
• Android Support by @AmSmart in #1130 / #1179
• Linux-ARM64 support by @nipeone in #1138 / #1179
• Feat/tensor override by @dpmm99 in #1180 and @martindevans in #1185
• Add LLamaReranker by @nipeone in #1150

Bug Fixes

• Adding extra values to the decode_result enum by @phil-scott-78 in #1140
• adding missing using statement by @NicklausBrain in #1154
• Fixed LLama.Web error starting session by @tonib in #1158
• Fixed CI (Ubuntu-24.04) by @martindevans in #1162
• Update LLamaEmbedder, Examples packages, and KernelMemory examples by @zsogitbe in #1170

Other Changes

• docs: Added link to new package MaIN.NET by @wisedev-pstach in #1155
• Create stale_issues.yml by @sangyuxiaowu in #1171
• docs: add deep-wiki link to readme. by @AsakusaRinne in #1172
• Update to M.E.AI 9.4.0-preview.1.25207.5 by @jozkee in #1156
• Update to M.E.AI 9.4.3-preview.1.25230.7 by @jozkee in #1182
• Update to M.E.AI 9.5.0-preview by @jozkee in #1186

New Contributors

• @NicklausBrain made their first contribution in #1154
• @wisedev-pstach made their first contribution in #1155
• @jozkee made their first contribution in #1156
• @tonib made their first contribution in #1158
• @nipeone made their first contribution in #1150

Full Changelog: v0.22.0...v0.24.0

v0.22.0

Major Changes

• Grammar Resampling by @martindevans, @m0nsky & @phil-scott-78 in #1109
• Added 'strict' parameter for template retrieval by @Lyrcaxis in #1124
• March 2025 version update by @martindevans in #1126

Bug Fixes

• Various fixes regarding special tokens (embeddings & anti-prompt) by @Lyrcaxis in #1112
• Fix Kernel Memory by @martindevans in #1081
• Updates json gbnf to match llama.cpp example by @phil-scott-78 in #1098
• InteractiveExecutor now stops at EOS tokens again by @dpmm99 in #1092

Other Changes

• Update README.md by @sangyuxiaowu in #1086
• Smaller Unit Test Model by @martindevans in #1105
• Update Microsoft.Extensions.AI to 9.3.0-preview.1.25114.11 by @stephentoub in #1102
• Cleaned up non-important warnings from the error list by @Lyrcaxis in #1111
• Update to M.E.AI 9.3.0-preview.1.25161.3 by @stephentoub in #1127
• Improved Memory Rental by @martindevans in #1113

Full Changelog: v0.21.0...v0.22.0

v0.21.0

Major Changes

• January 2025 Update by @martindevans in #1068

Full Changelog: v0.20.0...v0.21.0

v0.20.0

Major Changes

• January 2025 Update by @m0nsky, @Lyrcaxis, @SignalRT & @martindevans in #1036
• Removed .NET6 by @martindevans in #982
• Implemented LogitBias for DefaultSamplingPipeline by @martindevans in #986

Bug Fixes

• Fixed LLamaTokenDataArrayNative.Create's size initialization by @Lyrcaxis in #981
• Fixed ChatSession.LoadSession by @martindevans in #976
• Non-deterministic default seed by @Lyrcaxis in #993
• Fixed log levels (issue #995) by @LoicDagnas in #997
• Fix Batching shared token issues by @martindevans in #1045

Other Changes

• Adding simple batch example by @phil-scott-78 in #1038
• Add an option to apply the chat template to prompts when using StateslessExecutor by @phil-scott-78 in #1025
• Adds an exception for missing templates. by @phil-scott-78 in #1034
• Updated tests model to "LLama 3.2 1B Instruct" by @Lyrcaxis in #983
• Update M.E.AI.Abstractions by @stephentoub in #979, #992 and #1044
• Added sampling tests by @Lyrcaxis in #984
• Updated Nuget Dependencies by @martindevans in #987
• Helper method for sampling Conversation with SafeLLamaSamplerChainHandle by @martindevans in #994

New Contributors

• @LoicDagnas made their first contribution in #997
• @AcoranGonzalezMoray made their first contribution in #1015
• @phil-scott-78 made their first contribution in #1025

Full Changelog: v0.19.0...v0.20.0

v0.19.0

Major Changes

• November Binary Update by @martindevans in #962
• Custom Sampler Stages by @martindevans in #961
• Add Microsoft.Extensions.AI support for IChatClient / IEmbeddingGenerator by @stephentoub in #964

Bug Fixes

• Non-Null Default SamplingPipeline by @martindevans in #973

New Contributors

• @stephentoub made their first contribution in #964

Full Changelog: v0.18.0...v0.19.0

v0.18.0

Major Changes

• Split platform-specific binaries for NuGet backends by @m0nsky in #957

Other Changes

• Updates to ContributingGuide for latest llama.cpp repo by @scritch1sm in #953
• Fix README chat session example by @easis in #956

New Contributors

• @easis made their first contribution in #956

Full Changelog: v0.17.0...v0.18.0

v0.17.0

Important: The CUDA packages for 0.17.0 exceeded the maximum size for a nuget package. This means some of the 0.17.0 packages are not available until a new way is to deploy those packages. If you need one of the missing packages, use 0.16.0 instead.

Major Changes

• October binary update by @martindevans in #940

Bug Fixes

• Fix: Increase Vulkan detection timeout to 12 seconds by @LSXPrime in #931
• Specify file open permissions needed as read only. by @scritch1sm in #943
• Fix llama_get_logits_ith Null Handling by @martindevans in #920
• Fix CUDA compile error by @martindevans in #924

Other Changes

• Add LLama2 Chat Session example with a custom templator by @asmirnov82 in #938

New Contributors

• @LSXPrime made their first contribution in #931
• @scritch1sm made their first contribution in #943

Full Changelog: v0.16.0...v0.17.0