Tinfoil Release Notes

Qwen3-VL 30B (qwen3-vl-30b) has been removed from the catalog.

Added support for DeepSeek V4 Flash (the 0731 release), DeepSeek’s efficient MoE model with a 1M-token context window, built-in speculative decoding, tool calling, and structured outputs. Model ID: deepseek-v4-flash

Added support for Kimi K3, Moonshot’s flagship multimodal model for complex reasoning, long-horizon coding, visual knowledge work, and agentic workflows with tool calling.

Model ID: kimi-k3

Kimi K2.6 (kimi-k2-6) has been removed from the catalog.

DeepSeek V4 Pro (deepseek-v4-pro) has been removed from the catalog.

GLM-5.1 (glm-5-1) has been removed from the catalog. Use GLM-5.2 (glm-5-2) instead.

Added support for GLM-5.2, a frontier-class model built for agentic engineering with state-of-the-art performance on coding benchmarks and sustained reasoning over long tool-use sessions, with an extended 384K-token context window. Model ID: glm-5-2

Released the Tinfoil Rust SDK at v0.0.3 with attestation-verified TLS, an async-openai-compatible chat surface, and first-class shortcuts for chat (chat_relaxed), embeddings (embed_batch), and audio transcription (transcribe).

Streams Tinfoil-specific vendor events (web search progress, structured-outputs constraints, PII checks) through a permissive JSON builder.

Crate: tinfoilsh/tinfoil-rs

Whisper Large V3 Turbo (whisper-large-v3-turbo) has been removed from the catalog.

Use Voxtral Small 24B (voxtral-small-24b) for speech-to-text transcription.

PDF parsing moved from docling to sandboxed MuPDF; scanned pages route to the gemma4-31b VLM. Adds a mode query parameter on /v1/convert/file (text, vision, images, raw, vlm), multi-file uploads, and proper 502 failures on OCR errors. See the Document Processing Guide.

Added support for DeepSeek V4 Pro, a 1.6T-parameter Mixture-of-Experts model with 49B active parameters per token, strong long-context reasoning, coding, math, and agentic task performance.

Model ID: deepseek-v4-pro

Added support for Kimi K2.6, Moonshot’s latest multimodal agentic model with stronger long-horizon coding, design-to-code, image and video understanding, and large-scale tool-using workflows. Model ID: kimi-k2-6

Added support for GLM-5.1, a frontier-class model built for agentic engineering with state-of-the-art performance on coding benchmarks and sustained reasoning over long tool-use sessions. Model ID: glm-5-1

Added support for Gemma 4 31B, a multimodal model from Google DeepMind with image understanding, built-in thinking mode, and native function calling.

Model ID: gemma4-31b