v0.1.2-alpha

Third public alpha — a substantial agent-runtime and inference release.

Agent compatibility

• Add OpenAI Responses API and Anthropic Messages API support alongside Chat Completions.
• Stream tool-call events with strict schema-constrained arguments.
• Validate compatibility with Codex, Claude Code, Cline, OpenCode, Clu, and Ornith request shapes.
• Publish the agent-torture conformance and cross-runtime comparison tooling.

Runtime and performance

• Add continuous batched decoding and persistent/forkable KV prefixes.
• Share CUDA model weights across parallel slots and add optional Q8 KV cache.
• Fix CPU SIMD build flags and raise the default thread count for materially better out-of-box CPU throughput.
• Add optimized CUDA kernels while retaining token-identity gates.

Models and correctness

• Add native Ornith/Qwen3.5, Apertus, Gemma 4, SmolLM, Mistral, and Phi-3 tokenizer/template behavior.
• Strengthen strict JSON Schema support, including integer bounds and additional parser/schema OOM handling.
• Harden HTTP framing, graceful lifecycle handling, cross-process VRAM reservations, model swapping, and Python stream cancellation.

Release artifacts include Linux x86_64, macOS arm64, and Windows x86_64 binaries, packaged archives with provenance, and SHA-256 checksums.

Full diff: v0.1.1-alpha...v0.1.2-alpha

v0.1.1-alpha

Alpha stewardship release — everything the first wave of feedback asked for.

Compatibility

• OpenAI stop sequences: string or up to 4 strings, matched across token boundaries, excluded from output, in buffered and streamed responses.
• The AI-SDK agent-client dialect (Cline, OpenCode, …) works as-is: part-array message content, assistant tool_calls history, role:"tool" results, and stream_options.include_usage.

Hardening

• 10s request-read deadline: a stalled client gets a 408 and releases its inference slot instead of pinning it.
• Request-body allocation is null-checked (500 instead of a crash on OOM).

Project

• SECURITY.md documents the threat model (loopback-only, deliberate HTTP subset, model files as the untrusted input) and enables private vulnerability reporting.
• Issue forms require --version/--caps/model+quant; CONTRIBUTING states the correctness gates.

Full diff: v0.1.0-alpha...v0.1.1-alpha

v0.1.0-alpha

First public alpha of runner — a compact local LLM inference engine in plain C (GGUF, CPU/CUDA/Metal, OpenAI-compatible server, sampler-level JSON-schema enforcement).

Binaries: Linux x86_64 and Windows x86_64 (both need AVX2), macOS arm64 (Apple Silicon). Or build from source: make — no dependencies beyond a C compiler.

What to test: run your GGUF models on your hardware. If anything crashes, misbehaves, or underperforms, open an issue and include the output of runner --version and runner --caps.