Qualcomm GenieX CLI: baixar versions anteriors

Hexagon HTP
Microsoft-signed HTP catalog (llama.cpp @ e467e5). No cert import required on Windows on Snapdragon.
What's Changed
- feat(server): add GET /status endpoint by Mengsheng Wu (@mengshengwu) in #1560
Full Changelog: v0.9.1...v0.9.3

Hexagon HTP
Microsoft-signed HTP catalog (llama.cpp @ e467e5). No cert import required on Windows on Snapdragon.
What's Changed
- chore(sdk): bump geniex-qairt for FastRPC loader fixes by Perry Cheng (@zhycheng614) in #1556
- fix(sdk): join ADSP_LIBRARY_PATH with ';' on every platform by Perry Cheng (@zhycheng614) in #1555
Full Changelog: v0.9.0...v0.9.1

Hexagon HTP
Microsoft-signed HTP catalog (llama.cpp @ e467e5). No cert import required on Windows on Snapdragon.
What's Changed
- feat(sdk): add geniex_set_log_level and derive the QNN log level from it by David Qian (@Davidqian123) in #1510
- feat(bench): add --chat-template and default --log to error by David Qian (@Davidqian123) in #1513
- chore(sdk): bump geniex-qairt to latest main by David Qian (@Davidqian123) in #1516
- fix(sdk): fix HTP fastrpc mapping failures on IQ9/IQ8 NPU by Perry Cheng (@zhycheng614) in #1503
- feat(server): add Jev-shaped local decision endpoint by RemiliaForever (@RemiliaForever) in #1497
- fix(android): propagate apply_chat_template failures as exception by Gerard Louis Andres Recinto (@gerardrecinto) in #1518
- fix(sdk): stop pinning GPT-OSS experts to CPU by RemiliaForever (@RemiliaForever) in #1523
- chore(sdk): bump geniex-qairt for QAIRT tool-call support by RemiliaForever (@RemiliaForever) in #1525
- docs(docs): document reasoning_format for chat completions by RemiliaForever (@RemiliaForever) in #1528
- feat(server): expose OpenAI Decisions API by RemiliaForever (@RemiliaForever) in #1526
- fix(sdk): warm up llama.cpp context at load so HTP mapping stays out of TTFT by Mengsheng Wu (@mengshengwu) in #1529
- fix(sdk): enable flash attention on GPU only for GPT-OSS by zdc (@Hahahaooovvv) in #1524
- fix(sdk): avoid Windows Hexagon buffer cap by RemiliaForever (@RemiliaForever) in #1530
- feat(sdk): support caller-provided input embeddings for LLM generation by Mengsheng Wu (@mengshengwu) in #1482
- fix(sdk): run geniex-bench timing runs to n_gen tokens, ignoring EOS by Mengsheng Wu (@mengshengwu) in #1537
- feat(sdk): upgrade QAIRT plugin to 2.50 by RemiliaForever (@RemiliaForever) in #1535
- fix(sdk): resume partial range downloads safely by RemiliaForever (@RemiliaForever) in #1540
- docs(cli): add Windows driver update guide by RemiliaForever (@RemiliaForever) in #1541
- feat(sdk): geniex-bench self-abort on divergence by Perry Cheng (@zhycheng614) in #1517
- docs(cli): document VLM NPU workaround with a dedicated vision session by Mengsheng Wu (@mengshengwu) in #1546
- fix(sdk): attach media to geniex-bench VLM accuracy prompt by Mengsheng Wu (@mengshengwu) in #1543
- test(qdc): keep the device pytest session alive after a native abort by Mengsheng Wu (@mengshengwu) in #1539
- chore(sdk): bump llama.cpp submodule to e467e5b5e by Perry Cheng (@zhycheng614) in #1544
- fix(server): separate reasoning by default by RemiliaForever (@RemiliaForever) in #1545
- fix(sdk): map Android SM8750/SM8850 to for-Galaxy chipset aliases by Mengsheng Wu (@mengshengwu) in #1549
- feat(cli): add JEV structured decisions by Khoi VN (@vnk8071) in #1496
- ci(sdk): switch QDC artifact upload/log download to v2 API by Mengsheng Wu (@mengshengwu) in #1551
- fix(sdk): make llama.cpp OpenCL work on Adreno 663 (IQ-9075) by Mengsheng Wu (@mengshengwu) in #1538
- feat(sdk)!: add structured logits decisions by RemiliaForever (@RemiliaForever) in #1552
- fix(sdk): default HTP1 session so mmproj runs on its own HTP session by Perry Cheng (@zhycheng614) in #1548
New Contributors
- Gerard Louis Andres Recinto (@gerardrecinto) made their first contribution in #1518
- Khoi VN (@vnk8071) made their first contribution in #1496
Full Changelog: v0.8.0...v0.9.0

Hexagon HTP
Microsoft-signed HTP catalog (llama.cpp @ 942561). No cert import required on Windows on Snapdragon.
What's Changed
- docs(build): add instructions for building against different QAIRT SD… by Zack Li (@zhiyuan8) in #1500
- fix(sdk): put mmproj on a separate HTP session from the LM by Perry Cheng (@zhycheng614) in #1501
- chore(sdk): bump geniex-qairt to latest main by David Qian (@Davidqian123) in #1504
Full Changelog: v0.7.1...v0.8.0

Hexagon HTP
Microsoft-signed HTP catalog (llama.cpp @ 942561). No cert import required on Windows on Snapdragon.
What's Changed
- chore(sdk): bump llama.cpp submodule to c550d2f60 by Mengsheng Wu (@mengshengwu) in #1475
- chore(sdk): bump llama.cpp submodule to 94256114c by Mengsheng Wu (@mengshengwu) in #1483
- docs(docs): refresh README model sources and Android version by MAN$I VERMA 🦀 (@mansiverma897993) in #1445
- fix(sdk): default VLM vision encoder to FP16 instead of BF16 by Mengsheng Wu (@mengshengwu) in #1480
- feat(server): add prefill/decode t/s metrics to serve responses by Mengsheng Wu (@mengshengwu) in #1477
- fix(sdk): apply QCS6490 GPU batch defaults in geniex-bench by zdc (@Hahahaooovvv) in #1481
- fix(sdk): fix Adreno 643 Q4K OpenCL kernel compilation by zdc (@Hahahaooovvv) in #1484
- feat(sdk): add version-reporting parity across bindings, docker, and geniex-bench by Zack Li (@zhiyuan8) in #1491
- fix(ci): discover clang-cl.exe recursively for VS2026 Llvm layout by Perry Cheng (@zhycheng614) in #1492
- fix(sdk): always render special tokens when detokenizing llama_cpp output by RemiliaForever (@RemiliaForever) in #1472
- fix: parse MiniCPM5 XML tool calls by MAN$I VERMA 🦀 (@mansiverma897993) in #1462
Full Changelog: v0.7.0...v0.7.1

Hexagon HTP
Microsoft-signed HTP catalog (llama.cpp @ 4ff829). No cert import required on Windows on Snapdragon.
What's Changed
- fix(sdk): use resolved device for VLM encoder by RemiliaForever (@RemiliaForever) in #1437
- fix(sdk): patch llama.cpp OpenCL add kernel for Adreno by RemiliaForever (@RemiliaForever) in #1438
- ci: add Breeze grading workflow by RemiliaForever (@RemiliaForever) in #1424
- feat(qairt): enable HTP multicore for QAIRT bundles by MAN$I VERMA 🦀 (@mansiverma897993) in #1313
- feat(sdk): load QAIRT system prompt from metadata by RemiliaForever (@RemiliaForever) in #1443
- feat(sdk): expose the QAIRT runtime path on the C API by Mengsheng Wu (@mengshengwu) in #1441
- feat(sdk): add VLM compute device override by RemiliaForever (@RemiliaForever) in #1446
- fix(server): rebuild full VLM context after a failed generation by RemiliaForever (@RemiliaForever) in #1442
- chore(sdk): bump geniex-qairt to merged gemma4 VLM pipeline fix by David Qian (@Davidqian123) in #1448
- fix(sdk): keep bench output when generation exhausts the context by Perry Cheng (@zhycheng614) in #1451
- fix(server): decode base64 audio input by MAN$I VERMA 🦀 (@mansiverma897993) in #1458
- fix(go): retry chunk download on any error, not just timeout/EOF by Mengsheng Wu (@mengshengwu) in #1459
- fix(sdk): apply context params to VLM encoder by RemiliaForever (@RemiliaForever) in #1460
- feat(qairt): native stop sequences in the geniex-qairt pipeline by MAN$I VERMA 🦀 (@mansiverma897993) in #1342
- fix(sdk): defer geniex-bench top-p to bundle/plugin default when unset by Paul Zhu (@vinovo) in #1461
- fix(cli): print download cancel hint before progress by Mengsheng Wu (@mengshengwu) in #1468
- fix(qairt): surface QNN backend logs through --log/GENIEX_LOG by Mengsheng Wu (@mengshengwu) in #1470
- feat(sdk): add Qwen3.5-4B to geniex-bench model catalog by Mengsheng Wu (@mengshengwu) in #1469
- feat: unify HTP power-mode control across qairt and llama_cpp by Mengsheng Wu (@mengshengwu) in #1447
- fix(sdk): name supported formats when a repo has no loadable model files by zdc (@Hahahaooovvv) in #1467
Full Changelog: v0.6.1...v0.7.0

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.