v0.6.1

What's Changed

• fix: vlm does not have chat template applied in accuracy mode with pr… by Paul Zhu (@vinovo) in #1428

Full Changelog: v0.6.0...v0.6.1

Hexagon HTP

Microsoft-signed HTP catalog (llama.cpp @ 0eadef). No cert import required on Windows on Snapdragon.

v0.6.0

What's Changed

• docs(docs): geniex-bench tutorial for Windows and Linux by Mengsheng Wu (@mengshengwu) in #1371
• fix(server): honour deprecated max_tokens in chat completions by RemiliaForever (@RemiliaForever) in #1377
• fix(sdk): retry transient QDC 5xx during job status/log polling by Mengsheng Wu (@mengshengwu) in #1373
• fix(sdk): align llama_cpp MTP speculative decode with llama.cpp by Mengsheng Wu (@mengshengwu) in #1376
• fix(cli): release model resources when serve shuts down by Mengsheng Wu (@mengshengwu) in #1385
• feat(sdk): add geniex_model_get_detailed for single-model lookups by RemiliaForever (@RemiliaForever) in #1386
• chore(deps): bump geniex-qairt to fix KV-shift occupancy derivation by RemiliaForever (@RemiliaForever) in #1387
• fix(cli): restore the local precision picker for infer by RemiliaForever (@RemiliaForever) in #1381
• fix(cli): report a precision typo on infer instead of a generic SDK error by RemiliaForever (@RemiliaForever) in #1382
• fix(cli): hide already-downloaded precisions from the pull picker by RemiliaForever (@RemiliaForever) in #1383
• fix(sdk): keep sibling chunks alive when one chunk download fails by Mengsheng Wu (@mengshengwu) in #1388
• feat(cli): warn when a second Ctrl+C aborts serve's teardown by Mengsheng Wu (@mengshengwu) in #1392
• refactor(sdk): guard ggml backend dir scan behind GGML_BACKEND_DL by RemiliaForever (@RemiliaForever) in #1391
• fix(cli): match pre-run skip lists on the top-level command by RemiliaForever (@RemiliaForever) in #1394
• test(python): decode greedily and drive the device matrix from models.json by Mengsheng Wu (@mengshengwu) in #1384
• chore(deps): bump actions/setup-java from 5 to 6 by Dependabot (@dependabot)[bot] in #1399
• Fix/server model lookups by RemiliaForever (@RemiliaForever) in #1396
• fix(cli): harden the model-manager call surface by RemiliaForever (@RemiliaForever) in #1401
• fix(sdk): unblock the llama.cpp GPU path on RB3 Gen 2 by RemiliaForever (@RemiliaForever) in #1402
• chore: bump up geniex-qairt for smollm support by Paul Zhu (@vinovo) in #1404
• refactor(sdk): split geniex-bench into per-concern modules by RemiliaForever (@RemiliaForever) in #1378
• docs: document geniex model commands and RB3 Gen 2 VLM limit by RemiliaForever (@RemiliaForever) in #1406
• fix(sdk): default llama_cpp thread counts to llama.cpp's own value by Mengsheng Wu (@mengshengwu) in #1367
• ci(bench): add SmolVLM-2B and SmolVLM2-2.2B to the QDC bench matrix by Mengsheng Wu (@mengshengwu) in #1408
• feat(cli): stream text and tool calls as they complete by RemiliaForever (@RemiliaForever) in #1405
• feat(sdk): add preflight disk-space check to model pull by Mengsheng Wu (@mengshengwu) in #1416
• fix(sdk): local QAIRT precision fix + generic LLM family factory bump by David Qian (@Davidqian123) in #1419
• fix(server): keep tool messages in the chat completions prompt by RemiliaForever (@RemiliaForever) in #1407
• feat(cli): parse Qwen3.5, GPT-OSS and LFM2 tool calls by RemiliaForever (@RemiliaForever) in #1417
• fix(sdk): resolve untagged GGUF quants from general.file_type by Mengsheng Wu (@mengshengwu) in #1415
• fix(sdk): pin mtmd vision encoder to HTP for npu/hybrid by Mengsheng Wu (@mengshengwu) in #1420
• feat(sdk): apply chat template in geniex-bench --accuracy mode by Mengsheng Wu (@mengshengwu) in #1421
• fix(server): trigger keepalive reset from divergent conversations, not a header by RemiliaForever (@RemiliaForever) in #1422
• refactor(qairt): resolve the runtime path in one place, reachable from every binding by Perry Cheng (@zhycheng614) in #1423
• feat(sdk): re-introduce ModelScope hub for geniex pull by MAN$I VERMA (@mansiverma897993) in #1418
• feat(qairt): bundled QAIRT runtime by default, --qnn-lib as an optional override by Perry Cheng (@zhycheng614) in #1389
• chore(qairt): bump the plugin submodule to native-KV bundle support by David Qian (@Davidqian123) in #1425
• rename(cli): --qnn-lib -> --qairt-lib by Paul Zhu (@vinovo) in #1427

Full Changelog: v0.5.0...v0.6.0

Hexagon HTP

Microsoft-signed HTP catalog (llama.cpp @ 0eadef). No cert import required on Windows on Snapdragon.

v0.5.0

Hexagon HTP

Microsoft-signed HTP catalog (llama.cpp @ 873e5d). No cert import required on Windows on Snapdragon.

What's Changed

• ci(bench): add gemma-4 GPU cells and a --compute filter to QDC bench by Mengsheng Wu (@mengshengwu) in #1335
• fix(server): report speculative-decoding stats in OpenAI usage by Mengsheng Wu (@mengshengwu) in #1336
• ci(bench): add gemma-4-26B-A4B-it QAT GGUF for QDC bench by Mengsheng Wu (@mengshengwu) in #1347
• fix(bench): run VLM llama_cpp cells as text LLM in ctx/pp/tg sweep by Mengsheng Wu (@mengshengwu) in #1343
• fix(sdk): default llama.cpp swa_full to false by Mengsheng Wu (@mengshengwu) in #1346
• feat(server): parse Gemma 4 tool-call format by RemiliaForever (@RemiliaForever) in #1345
• feat(sdk): add MTP speculative-decoding coverage to bench matrix by Mengsheng Wu (@mengshengwu) in #1348
• feat(sdk): return PROMPT_TOO_LONG when the prompt itself overflows by RemiliaForever (@RemiliaForever) in #1349
• feat(sdk): measure media encoder time separately from prefill by RemiliaForever (@RemiliaForever) in #1354
• fix(cli): drop unsupported modalities before VLM generate by RemiliaForever (@RemiliaForever) in #1356
• fix(cli): run update installer silently by MAN$I VERMA (@mansiverma897993) in #1352
• test(sdk): add CPU + qairt unit tests for chat template, KV shift, mtmd by Mengsheng Wu (@mengshengwu) in #1355
• fix(bench): raise QDC job timeout and drop hybrid from the matrix by Mengsheng Wu (@mengshengwu) in #1358
• test(sdk): add cross-backend logits parity check by Mengsheng Wu (@mengshengwu) in #1357
• test(qdc): skip test_mtp_multi_turn on QCS9075M by Mengsheng Wu (@mengshengwu) in #1359
• feat(sdk): inject tools into the QAIRT VLM chat template by RemiliaForever (@RemiliaForever) in #1360
• feat(sdk): accept an explicit device list as a compute unit by Mengsheng Wu (@mengshengwu) in #1361
• fix(sdk): reuse VLM KV via char-level prefix match by RemiliaForever (@RemiliaForever) in #1363
• test(python): disable thinking in llm quality keyword cells by RemiliaForever (@RemiliaForever) in #1365
• fix(resume): re-fetch file when size mismatches legacy no-marker state by Mengsheng Wu (@mengshengwu) in #1364
• feat(sdk): add a CPU-only linux build variant for baseline armv8.0 boards by David Qian (@Davidqian123) in #1340
• feat(sdk): resolve AIHM version from latest.txt instead of pinning by zdc (@Hahahaooovvv) in #1353
• test(sdk): stop leaking tempdirs in model-manager store tests by RemiliaForever (@RemiliaForever) in #1368
• feat(android): add a CPU-only AAR for baseline armv8.0 devices by RemiliaForever (@RemiliaForever) in #1366
• docs(docs): document multi-HTP device list in MTP tutorial and CLI reference by Mengsheng Wu (@mengshengwu) in #1370
• feat(sdk): fetch AIHM index from releases/latest/, pinned as fallback by zdc (@Hahahaooovvv) in #1372
• chore: bump up llama.cpp by Paul Zhu (@vinovo) in #1374

Full Changelog: v0.4.0...v0.5.0

v0.4.0

Hexagon HTP

Microsoft-signed HTP catalog (llama.cpp @ 6ba5ef). No cert import required on Windows on Snapdragon.

What's Changed

• chore(deps): bump docker/login-action from 4.5.2 to 4.6.0 by Dependabot (@dependabot)[bot] in #1270
• docs(release): add release check list and fix HTP file count by RemiliaForever (@RemiliaForever) in #1265
• feat(cli/server): hint --host 0.0.0.0 when bound to loopback by RemiliaForever (@RemiliaForever) in #1264
• fix(sdk): restore mtmd_input_text.text_len in VLM tokenize by MAN$I VERMA (@mansiverma897993) in #1273
• fix(sdk): select clip backend from the resolved device, not ngl by Sanoop Sadheerthan (@sanoop-sadheerthan) in #1248
• feat(cli): show models cache root on list by RemiliaForever (@RemiliaForever) in #1263
• feat(cli): show size/location/precision on pull by RemiliaForever (@RemiliaForever) in #1271
• feat(cli): print backend/device summary on infer --verbose by RemiliaForever (@RemiliaForever) in #1274
• fix(go): print only crashing goroutine on backend abort() by RemiliaForever (@RemiliaForever) in #1275
• fix(cli/server): print qairt compute-coercion warning on serve (#1295) by RemiliaForever (@RemiliaForever) in #1276
• docs(docs): add tutorials section with MTP guide and explicit chipset list by Alan Zhu (@alanzhuly) in #1280
• fix(sdk): filter chipset list to the host OS by RemiliaForever (@RemiliaForever) in #1279
• geniex-bench: honour --prompt-file in the VLM run loop by Paul Zhu (@vinovo) in #1278
• docs: document --nctx and improve context-exceeded UX by RemiliaForever (@RemiliaForever) in #1281
• feat(cli): add --log flag equivalent to GENIEX_LOG by RemiliaForever (@RemiliaForever) in #1283
• docs(cli): refresh update-checks section in run.md by RemiliaForever (@RemiliaForever) in #1287
• feat(python): show progress bar during pip install SDK fetch by Mengsheng Wu (@mengshengwu) in #1282
• perf(python): coalesce range GETs in install-time SDK fetch by Mengsheng Wu (@mengshengwu) in #1288
• build(release): rename Python sdist assets to match other release artifacts by Mengsheng Wu (@mengshengwu) in #1290
• ci(qdc): reconstruct pytest summary from --report-log on abort by Mengsheng Wu (@mengshengwu) in #1291
• feat(sdk): add InternVL 3.5 QAIRT model support and fix Gemma4 geniex-bench errors by David Qian (@Davidqian123) in #1285
• fix(sdk): reacquire HTP sessions on all devices, not just NPU by Mengsheng Wu (@mengshengwu) in #1289
• fix(cli): echo compute alias in model-loaded line by RemiliaForever (@RemiliaForever) in #1296
• fix(sdk): patch llama.cpp OpenCL cpy kernel for Adreno by RemiliaForever (@RemiliaForever) in #1284
• fix(server): send spec-compliant finish_reason on stream chunks by MAN$I VERMA (@mansiverma897993) in #1246
• feat(server): match tool-call JSON by braces instead of tags by RemiliaForever (@RemiliaForever) in #1298
• fix(cli): apply --log to the SDK log callback by RemiliaForever (@RemiliaForever) in #1299
• fix(cli): suggest --log debug in model-load failure hint by RemiliaForever (@RemiliaForever) in #1300
• ci: split test surface into Unit Test + QDC Test workflows by Mengsheng Wu (@mengshengwu) in #1297
• fix(ci): avoid rebuilding SDK inside QDC Test when called from test.yml by Mengsheng Wu (@mengshengwu) in #1302
• fix(sdk): fully clear VLM memory on reset by zdc (@Hahahaooovvv) in #1301
• fix(cli): route --spec-type through the LLM path for VLM-classified models by Mengsheng Wu (@mengshengwu) in #1295
• ci(test): change-gate PR runs, tag-gate full matrix, JSON model manifest by Mengsheng Wu (@mengshengwu) in #1303
• feat(cli): add 'geniex model list' for AI Hub catalogue by RemiliaForever (@RemiliaForever) in #1304
• ci(qdc): route matrix values through env to satisfy zizmor by Mengsheng Wu (@mengshengwu) in #1307
• ci(test): add Rust model-manager unit-test job to PR check by Mengsheng Wu (@mengshengwu) in #1309
• feat(cli): auto-detect RB3 Gen2 and default compute to cpu by RemiliaForever (@RemiliaForever) in #1306
• chore: bump up geniex-qairt to include qnwe3 0.6b and gemma 4 fix by Paul Zhu (@vinovo) in #1310
• test(qdc): run llama_cpp LLM + VLM QDC matrix on cpu/gpu/npu by Mengsheng Wu (@mengshengwu) in #1308
• fix(server): route reasoning_format to split thinking into reasoning_content by RemiliaForever (@RemiliaForever) in #1305
• refactor(server): split chat handler and unify token routing by RemiliaForever (@RemiliaForever) in #1315
• refactor(sdk)!: remove unused fields from geniex.h and sync bindings by RemiliaForever (@RemiliaForever) in #1319
• docs(docs): use PowerShell backtick continuation in mtp tutorial by Mengsheng Wu (@mengshengwu) in #1320
• docs(docs): trim mtp tutorial and flatten tutorials nav by Alan Zhu (@alanzhuly) in #1321
• feat(server): add raw /v1/completions for FIM autocompletion by MAN$I VERMA (@mansiverma897993) in #1317
• Add Android SM8850 leg to QDC test matrix by Mengsheng Wu (@mengshengwu) in #1318
• docs(docs): document audio input across CLI, server, and Python by RemiliaForever (@RemiliaForever) in #1325
• sdk(model-manager): bump DEFAULT_AI_HUB_VERSION to v0.60.0 by David Qian (@Davidqian123) in #1327
• fix(sdk): only last MTP-prefill token needs logits by Mengsheng Wu (@mengshengwu) in #1333
• fix(sdk): fix MTP failures caused by HTP session churn and VLM model leak by Mengsheng Wu (@mengshengwu) in #1334
• fix(server): guard in-use models from keep-alive destruction by MAN$I VERMA (@mansiverma897993) in #1324
• fix(sdk): accept both QAIRT runtime enums when selecting AI Hub assets by David Qian (@Davidqian123) in #1338

New Contributors

• Sanoop Sadheerthan (@sanoop-sadheerthan) made their first contribution in #1248
• zdc (@Hahahaooovvv) made their first contribution in #1301

Full Changelog: v0.3.18...v0.4.0

v0.3.20

Hexagon HTP

Microsoft-signed HTP catalog (llama.cpp @ 6ba5ef). No cert import required on Windows on Snapdragon.

What's Changed

• test(qdc): run llama_cpp LLM + VLM QDC matrix on cpu/gpu/npu by Mengsheng Wu (@mengshengwu) in #1308
• fix(server): route reasoning_format to split thinking into reasoning_content by RemiliaForever (@RemiliaForever) in #1305
• refactor(server): split chat handler and unify token routing by RemiliaForever (@RemiliaForever) in #1315
• refactor(sdk)!: remove unused fields from geniex.h and sync bindings by RemiliaForever (@RemiliaForever) in #1319
• docs(docs): use PowerShell backtick continuation in mtp tutorial by Mengsheng Wu (@mengshengwu) in #1320
• docs(docs): trim mtp tutorial and flatten tutorials nav by Alan Zhu (@alanzhuly) in #1321
• feat(server): add raw /v1/completions for FIM autocompletion by MAN$I VERMA (@mansiverma897993) in #1317
• Add Android SM8850 leg to QDC test matrix by Mengsheng Wu (@mengshengwu) in #1318
• docs(docs): document audio input across CLI, server, and Python by RemiliaForever (@RemiliaForever) in #1325
• sdk(model-manager): bump DEFAULT_AI_HUB_VERSION to v0.60.0 by David Qian (@Davidqian123) in #1327
• fix(sdk): only last MTP-prefill token needs logits by Mengsheng Wu (@mengshengwu) in #1333
• fix(sdk): fix MTP failures caused by HTP session churn and VLM model leak by Mengsheng Wu (@mengshengwu) in #1334
• fix(server): guard in-use models from keep-alive destruction by MAN$I VERMA (@mansiverma897993) in #1324

Full Changelog: v0.3.19...v0.3.20

v0.3.19

What's Changed

• chore(deps): bump docker/login-action from 4.5.2 to 4.6.0 by Dependabot (@dependabot)[bot] in #1270
• docs(release): add release check list and fix HTP file count by RemiliaForever (@RemiliaForever) in #1265
• feat(cli/server): hint --host 0.0.0.0 when bound to loopback by RemiliaForever (@RemiliaForever) in #1264
• fix(sdk): restore mtmd_input_text.text_len in VLM tokenize by MAN$I VERMA (@mansiverma897993) in #1273
• fix(sdk): select clip backend from the resolved device, not ngl by Sanoop Sadheerthan (@sanoop-sadheerthan) in #1248
• feat(cli): show models cache root on list by RemiliaForever (@RemiliaForever) in #1263
• feat(cli): show size/location/precision on pull by RemiliaForever (@RemiliaForever) in #1271
• feat(cli): print backend/device summary on infer --verbose by RemiliaForever (@RemiliaForever) in #1274
• fix(go): print only crashing goroutine on backend abort() by RemiliaForever (@RemiliaForever) in #1275
• fix(cli/server): print qairt compute-coercion warning on serve (#1295) by RemiliaForever (@RemiliaForever) in #1276
• docs(docs): add tutorials section with MTP guide and explicit chipset list by Alan Zhu (@alanzhuly) in #1280
• fix(sdk): filter chipset list to the host OS by RemiliaForever (@RemiliaForever) in #1279
• geniex-bench: honour --prompt-file in the VLM run loop by Paul Zhu (@vinovo) in #1278
• docs: document --nctx and improve context-exceeded UX by RemiliaForever (@RemiliaForever) in #1281
• feat(cli): add --log flag equivalent to GENIEX_LOG by RemiliaForever (@RemiliaForever) in #1283
• docs(cli): refresh update-checks section in run.md by RemiliaForever (@RemiliaForever) in #1287
• feat(python): show progress bar during pip install SDK fetch by Mengsheng Wu (@mengshengwu) in #1282
• perf(python): coalesce range GETs in install-time SDK fetch by Mengsheng Wu (@mengshengwu) in #1288
• build(release): rename Python sdist assets to match other release artifacts by Mengsheng Wu (@mengshengwu) in #1290
• ci(qdc): reconstruct pytest summary from --report-log on abort by Mengsheng Wu (@mengshengwu) in #1291
• feat(sdk): add InternVL 3.5 QAIRT model support and fix Gemma4 geniex-bench errors by David Qian (@Davidqian123) in #1285
• fix(sdk): reacquire HTP sessions on all devices, not just NPU by Mengsheng Wu (@mengshengwu) in #1289
• fix(cli): echo compute alias in model-loaded line by RemiliaForever (@RemiliaForever) in #1296
• fix(sdk): patch llama.cpp OpenCL cpy kernel for Adreno by RemiliaForever (@RemiliaForever) in #1284
• fix(server): send spec-compliant finish_reason on stream chunks by MAN$I VERMA (@mansiverma897993) in #1246
• feat(server): match tool-call JSON by braces instead of tags by RemiliaForever (@RemiliaForever) in #1298
• fix(cli): apply --log to the SDK log callback by RemiliaForever (@RemiliaForever) in #1299
• fix(cli): suggest --log debug in model-load failure hint by RemiliaForever (@RemiliaForever) in #1300
• ci: split test surface into Unit Test + QDC Test workflows by Mengsheng Wu (@mengshengwu) in #1297
• fix(ci): avoid rebuilding SDK inside QDC Test when called from test.yml by Mengsheng Wu (@mengshengwu) in #1302
• fix(sdk): fully clear VLM memory on reset by zdc (@Hahahaooovvv) in #1301
• fix(cli): route --spec-type through the LLM path for VLM-classified models by Mengsheng Wu (@mengshengwu) in #1295
• ci(test): change-gate PR runs, tag-gate full matrix, JSON model manifest by Mengsheng Wu (@mengshengwu) in #1303
• feat(cli): add 'geniex model list' for AI Hub catalogue by RemiliaForever (@RemiliaForever) in #1304
• ci(qdc): route matrix values through env to satisfy zizmor by Mengsheng Wu (@mengshengwu) in #1307
• ci(test): add Rust model-manager unit-test job to PR check by Mengsheng Wu (@mengshengwu) in #1309
• feat(cli): auto-detect RB3 Gen2 and default compute to cpu by RemiliaForever (@RemiliaForever) in #1306
• chore: bump up geniex-qairt to include qnwe3 0.6b and gemma 4 fix by Paul Zhu (@vinovo) in #1310

New Contributors

• Sanoop Sadheerthan (@sanoop-sadheerthan) made their first contribution in #1248
• zdc (@Hahahaooovvv) made their first contribution in #1301

Full Changelog: v0.3.18...v0.3.19

Hexagon HTP

Microsoft-signed HTP catalog (llama.cpp @ 6ba5ef). No cert import required on Windows on Snapdragon.

v0.3.18

Hexagon HTP

Support Gemma4-E2B and Gemma4-E4B QAIRT inference through local path.
Microsoft-signed HTP catalog (llama.cpp @ 6ba5ef). No cert import required on Windows on Snapdragon.

What's Changed

• docs(install): cross-link Docker and native Linux install paths by Mengsheng Wu (@mengshengwu) in #1231
• docs(server): rewrite tool-calling example as VLM + web_search round-trip by Mengsheng Wu (@mengshengwu) in #1238
• fix(sdk): forward quant suffix on benchmark cache hit by RemiliaForever (@RemiliaForever) in #1239
• ci: harden workflows against zizmor high-severity findings by RemiliaForever (@RemiliaForever) in #1240
• chore(release): source signed HTP bundle from qcom-ai-hub/geniex LFS by Mengsheng Wu (@mengshengwu) in #1237
• chore(deps): bump actions/stale from 10 to 11 by Dependabot (@dependabot)[bot] in #1249
• chore(sdk): drop stale gitattributes and gitmodules by Mengsheng Wu (@mengshengwu) in #1251
• chore(deps): bump pypa/gh-action-pypi-publish from 1.14.1 to 1.14.2 by Dependabot (@dependabot)[bot] in #1253
• refactor(sdk): drop default-quant from model-manager by Mengsheng Wu (@mengshengwu) in #1252
• feat(sdk)!: generalize llama_cpp speculative decoding to all types by Mengsheng Wu (@mengshengwu) in #1195
• feat(qairt): Add Gemma 4 QAIRT support by David Qian (@Davidqian123) in #1257
• chore(deps): bump docker/login-action from 4 to 4.5.2 by Dependabot (@dependabot)[bot] in #1258
• fix(sdk): key QAIRT ModelFile by precision so name:precision ids resolve by Mengsheng Wu (@mengshengwu) in #1261
• fix(sdk): recognize scalar ggml quant tags (F/BF/FP/I) by Mengsheng Wu (@mengshengwu) in #1262
• feat(sdk): expose raw logits (geniex_llm_forward_logits) in bench + serve by RemiliaForever (@RemiliaForever) in #1256

Full Changelog: v0.3.17...v0.3.18

v0.3.17

Hexagon HTP

Microsoft-signed HTP catalog (llama.cpp @ ae9291). No cert import required on Windows on Snapdragon.

What's Changed

• chore(deps): bump actions/setup-python from 6 to 7 by Dependabot (@dependabot)[bot] in #1205
• fix(cli): exclude resumed bytes from download speed by Mengsheng Wu (@mengshengwu) in #1206
• ci(release): use Trusted Publishing for TestPyPI by Mengsheng Wu (@mengshengwu) in #1207
• ci: remove manual llama.cpp Windows ARM64 build workflow by Mengsheng Wu (@mengshengwu) in #1194
• docs(cli): move host-dependency install into CLI page by RemiliaForever (@RemiliaForever) in #1213
• docs(cli): tidy host-dependency intro wording by RemiliaForever (@RemiliaForever) in #1214
• docs(cli): add libqnn1 for linux by RemiliaForever (@RemiliaForever) in #1215
• docs(server): add non-streaming and tool-calling examples by Mengsheng Wu (@mengshengwu) in #1208
• fix(server): populate id and type on tool_calls response by Mengsheng Wu (@mengshengwu) in #1212
• fix(sdk): fold :default quant to its lower-case sentinel by MAN$I VERMA (@mansiverma897993) in #1216
• fix(server): unblock cross-origin browser clients in CORS middleware by MAN$I VERMA (@mansiverma897993) in #1222
• feat(sdk): auto-detect Dragonwing IQ-8275 EVK via socinfo by Mengsheng Wu (@mengshengwu) in #1224
• feat(sdk): honor HF_ENDPOINT for HuggingFace downloads by Mengsheng Wu (@mengshengwu) in #1221
• feat(cli): gate windows update on installer signing by RemiliaForever (@RemiliaForever) in #1219
• chore(sdk): bump pinned aihub release to v0.58.0 by RemiliaForever (@RemiliaForever) in #1226
• feat(sdk): auto-detect Snapdragon X2 Plus hosts by MAN$I VERMA (@mansiverma897993) in #1228
• feat(sdk): detect chipset via Adreno GPU by CZ Zhang (@changzheng-zhang) in #1229
• chore(sdk): trim excess comments in detect.rs by Mengsheng Wu (@mengshengwu) in #1230
• docs: bump toolchain image to v0.1.0 and update release notes by RemiliaForever (@RemiliaForever) in #1232
• fix(sdk): invalidate aihub index cache on version change by RemiliaForever (@RemiliaForever) in #1234

New Contributors

• MAN$I VERMA (@mansiverma897993) made their first contribution in #1216
• CZ Zhang (@changzheng-zhang) made their first contribution in #1229

Full Changelog: v0.3.16...v0.3.17

v0.3.16

Hexagon HTP

Microsoft-signed HTP catalog (llama.cpp @ ae9291). No cert import required on Windows on Snapdragon.

What's Changed

• docs: simplify Docker install to Docker Hub only by Zack Li (@zhiyuan8) in #1139
• docs(dx): add same NPU troubleshooting to Chinese docs by Perry Cheng (@zhycheng614) in #1142
• feat(server): add server-wide nctx/ngl/compute defaults to serve by RemiliaForever (@RemiliaForever) in #1141
• fix(ci): unify source license headers to BSD-3-Clause by RemiliaForever (@RemiliaForever) in #1144
• docs: drop HOME export step from Linux ARM64 install by Alan Zhu (@alanzhuly) in #1145
• fix(python): fix VLM multi-turn image guard and test message order by Mengsheng Wu (@mengshengwu) in #1130
• fix(sdk): report qairt prefill rate over padded prompt length by RemiliaForever (@RemiliaForever) in #1148
• fix(sdk): load non-ASCII image paths in qairt VLM on Windows by RemiliaForever (@RemiliaForever) in #1152
• docs: add Simplified Chinese README and sync cn docs by RichardDoong in #1146
• chore(ci): drop dead .apk content-type branch in release.js by RemiliaForever (@RemiliaForever) in #1158
• docs(python): update install instructions to use production PyPI by Mengsheng Wu (@mengshengwu) in #1151
• feat(sdk): support input_ids and get_model_info in qairt plugin by David Qian (@Davidqian123) in #1157
• ci(release): use Trusted Publishing for production PyPI by Perry Cheng (@zhycheng614) in #1156
• feat(sdk): support batched prompts in geniex-bench via --- separator by RemiliaForever (@RemiliaForever) in #1159
• docs(linux): surface host-dependency setup as an install pre-step by Mengsheng Wu (@mengshengwu) in #1167
• feat(qairt): opt-in sliding window CLI flag and env var by David Qian (@Davidqian123) in #1160
• ci(build-sdk): build unsigned HTP skels when signing secret is absent by Mengsheng Wu (@mengshengwu) in #1171
• fix(cli): align serve short aliases with infer/run by RemiliaForever (@RemiliaForever) in #1172
• ci(build-sdk): override cached HTP cert path so unsigned fork builds skip signing by Mengsheng Wu (@mengshengwu) in #1173
• feat(sdk): pull models from Docker Hub's ai/* namespace by Eric Curtin (@ericcurtin) in #1165
• fix(sdk): default to npu and stop clobbering explicit ngl by RemiliaForever (@RemiliaForever) in #1174
• feat(cli): fetch updates from S3 index instead of GitHub releases by RemiliaForever (@RemiliaForever) in #1164
• chore(deps): bump actions/setup-node from 6 to 7 by Dependabot (@dependabot)[bot] in #1176
• fix(sdk): fail cleanly on CPUs missing required armv8.2 features by RemiliaForever (@RemiliaForever) in #1180
• chore(sdk): bump geniex-qairt submodule pointer by David Qian (@Davidqian123) in #1175
• refactor(sdk): remove dead license_key / validation subsystem by Mengsheng Wu (@mengshengwu) in #1184
• chore(sdk): bump llama.cpp to b10019 by RemiliaForever (@RemiliaForever) in #1182
• Remove Mixpanel integration from documentation by AlexCHEN (@alexchen4ai) in #1188
• chore(deps): bump actions/setup-go from 6 to 7 by Dependabot (@dependabot)[bot] in #1189
• docs(docs): expand contributing guide for open-source contributors by Mengsheng Wu (@mengshengwu) in #1177
• docs(cli): sync Chinese README with English CLI docs by Ting Heng Nguan (@HengNguan) in #1192
• fix(ci): bump Windows Hexagon SDK to 6.6.0.0 for parity with Linux/Android by Mengsheng Wu (@mengshengwu) in #1193
• fix(python): exclude resumed bytes from download speed by Mengsheng Wu (@mengshengwu) in #1190
• fix(sdk): fix gemma-4-E2B VLM load and MTP quant selection by RemiliaForever (@RemiliaForever) in #1197
• fix(sdk): enable clip GPU backend when n_gpu_layers is -1 by RemiliaForever (@RemiliaForever) in #1198
• chore(deps): bump pypa/gh-action-pypi-publish from 1.14.0 to 1.14.1 by Dependabot (@dependabot)[bot] in #1200

New Contributors

• RichardDoong made their first contribution in #1146
• Eric Curtin (@ericcurtin) made their first contribution in #1165
• Ting Heng Nguan (@HengNguan) made their first contribution in #1192

Full Changelog: v0.3.14...v0.3.16

v0.3.14

Hexagon HTP

Microsoft-signed HTP catalog (llama.cpp @ be4a6a). No cert import required on Windows on Snapdragon.

What's Changed

• chore(deps): bump GitHub Actions (checkout, setup-buildx-action, configure-pages) by RemiliaForever (@RemiliaForever) in #1123
• fix(sdk): auto-detect QAIRT VLM bundles in geniex-bench by David Qian (@Davidqian123) in #1124
• Added developer preview tag by Alan Zhu (@alanzhuly) in #1127
• fix(build): bump Hexagon toolchain to v0.7 to fix miscompiled HTP v75… by Perry Cheng (@zhycheng614) in #1128
• fix(docs): update style.css by RemiliaForever (@RemiliaForever) in #1129
• chore(sdk): bump pinned aihm release to v0.57.0 by David Qian (@Davidqian123) in #1133
• chore: add BSD-3-Clause SPDX headers to satisfy Repolinter by AlexCHEN (@alexchen4ai) in #1136
• ci(release): publish stable tags to production PyPI by Paul Zhu (@vinovo) in #1134
• docs(dx): add NPU driver symlink and hybrid CPU-fallback troubleshooting by Perry Cheng (@zhycheng614) in #1137

Full Changelog: v0.3.13...v0.3.14