Tencent releases Hy4 flash as open weights on Hugging Face, plus 1 more update
Version 1
In Brief
Details
Tencent — Hy4 flash is now open-weight
Tencent
Tencent released Hy4 flash, Hy4 flash-FP8 and the pre-trained Hy4 flash-base as open-weight models on Hugging Face, ModelScope and GitCode.
- Tencent released three open-weight checkpoints: the Hy4 flash instruct model, Hy4 flash-FP8 (FP8-quantized instruct model) and Hy4 flash-base (pre-trained base model, not instruction-tuned), on Hugging Face, ModelScope and GitCode.
- The weights are released under the Tencent Hy License; code and documentation are in the Tencent-Hunyuan/Hy4-flash GitHub repository.
- Per the model card, Hy4 flash has 220B total parameters (including 63.5B Engram parameters) with about 12B active per token, accepts text and image input, generates text, and supports a context of up to 1,048,576 tokens.
- The model card lists 42 backbone layers (1 dense, 41 MoE), 512 routed experts plus 1 shared expert per MoE layer with 16 routed experts active per token, and a built-in 0.85B DFly2 drafter for speculative decoding.
- Deployment instructions cover Transformers (support to be merged into main soon; install from a specific commit until then), an official vLLM image (vllm/vllm-openai:hy4-flash) and an official SGLang image; reasoning_effort supports "high" (default) and "no_think".
Sources:Hugging Face model card✓ officialGitHub✓ officialTencent Hy✓ officialTencent Hunyuan on X✓ official
OpenAI — Codex 0.171.0-alpha.13 adds TUI remote control and subagent sidebar views
OpenAI
Codex 0.171.0-alpha.13, a GitHub pre-release, adds a /rc remote-control command and subagent views in the TUI, device-code login for config servers, and GPT-6.2 Sol and Astra in model catalogs.
- OpenAI published Codex 0.171.0-alpha.13 as a pre-release on GitHub under the rust-v0.171.0-alpha.13 tag.
- The TUI gains a /rc command to manage remote control, a keyboard diagram in shortcut help, and subagent status in the sidebar chooser, with subagent transcripts openable from the sidebar and from activity links.
- Device-code login is supported for config servers, and enterprise MCP sign-in works without refresh tokens.
- GPT-6.2 Sol and Astra were added to model catalogs, the spawn_agent model listing cap was expanded to 16, and audio transcription was added to the backend client.
- Thread search can stream partial results and support cancellation, and the app-server socket relay was moved into the Codex CLI.
- Fixes include escaping installer PATH entries for shell metacharacters, Node launcher detection in the standalone installer, stray input when scrolling (via a crossterm update), keeping the TUI composer caret anchored while typing, and repairing missing or stale Windows sandbox Core registration.
Sources:Codex 0.171.0-alpha.13 release✓ officialGitHub compare versus 0.171.0-alpha.12✓ official