Release 0.57.0

[michalharakal]

Release 0.57.0, in lock-step with SKaiNET engine 0.57.0.

Contents

• Qwen on the compiled IREE KV path (Feature/qwen iree kv export #457, closes Qwen export harness (StableHLO + parameters) for the IREE path — none exists; FunctionGemma is the only tool-calling model with a compiled leg #411, llm-agent: QwenChatTemplate(enableThinking = false) pre-fills an empty <think> block, which Qwen2.5-Instruct never saw — tool calling drops to 0–1/8 vs 4/8 with the official Qwen2.5 template #409). The release includes the qwen-kv-v1 export (QwenExportHarness / QwenExportCli), the grouped-query IreeKvSession with rebuilt libskainet_iree_kv.so for both ABIs, and Qwen25ChatTemplate.
  • Qwen3-0.6B compiles for Vulkan valhall4 and arm32 and matches llama.cpp's 32 greedy tokens on host IREE (QwenVmfbParityTest).
  • It has not been run on a device yet.
  • Qwen2.5 export is blocked by the core attention-bias gap.
• Moonshine v2 stream: the final re-decode is capped at the model card's length budget (moonshine stream: cap the final decode budget at the model card's output length #455).
• Engine 0.57.0 and Kotlin 2.4.20 (chore: consume engine 0.57.0; Kotlin 2.4.20 #453), the FunctionGemma embedding geometry in the export contract (functiongemma: carry the embedding geometry in the export contract #452), and dependency bumps: Ktor 3.6.0 (build(deps): Bump ktorClientCore from 3.5.2 to 3.6.0 #448), kctfork 0.14.0 (build(deps): Bump kctfork from 0.13.0 to 0.14.0 #449), jackson-databind 2.22.3 (build(deps): Bump com.fasterxml.jackson.core:jackson-databind from 2.22.2 to 2.22.3 #454).

Release preparation

• CHANGELOG 0.57.0 entry, README "Current release" section, docs coordinates, and VERSION_NAME.
• Merge develop into release/0.57.0 brings in Feature/qwen iree kv export #457, moonshine stream: cap the final decode budget at the model card's output length #455 and the dependency bumps. The CHANGELOG conflict is resolved into the single 0.57.0 section.

Verified locally after the merge, all green:

• apiCheck.
• jvmTest for llm-core, qwen, functiongemma and llm-agent.
• The iree-android unit tests.

Gates

• Publish dry run on release/0.57.0: https://github.com/SKaiNET-developers/SKaiNET-transformers/actions/runs/36154797077
• After merge (with a merge commit): wait for develop CI on the merge commit, tag it 0.57.0, and wait for the publish run.