v7.9.0

AX Engine v7.9.0

AX Engine 7.9.0 expands chat compatibility and constrained generation, improves Qwen execution paths, and builds with pinned Rust 1.99.1. Rust, Python, JavaScript, Go, Ruby, and Swift version surfaces advance together.

• Release an idle upstream generation stream as soon as an Anthropic streaming client disconnects.
• Add Anthropic Messages streaming, tool use, thinking, image-block translation, and native prompt token counting over the existing chat path.
• Reject malformed UTF-8 and unpaired Unicode surrogates during constrained decoding; count escaped surrogate pairs as one character for string-length limits.
• Constrain JSON response generation with the supported JSON Schema subset, including integration with speculative decoding and final response validation.
• Apply OpenAI presence and frequency penalties and logit bias, and restore the Qwen pre-tokenizer expression used by Qwen 3.5-generation packs.
• Improve hybrid-model prefix reuse, Flash Next n-gram reads and attention-cache storage, and native decode operations.
• Add the Qwen 3.8 27B DFlash2 block drafter path with model-specific admission and adaptive verification windows.
• Use the Tiel tool template for MXFP8 packs and return actionable Hub download failures while restricting download credentials to the official Hub.
• Update TLS certificate dependencies and pin the build toolchain to Rust 1.99.1.

Model qualification and performance claims retain their documented hardware, pack, execution-mode, and evidence scope. This release does not establish new certification or a general performance multiplier.

Full Changelog: v7.8.0...v7.9.0