#open source

Tencent has released Hy4, an open-weight foundation model, with open-sourced HyCell and HyPDE models for single-cell representation and physical-field modeling. For game development, Tencent says it worked with its Interactive Entertainment Group to curate training data and evaluations covering feature implementation, gameplay logic, state tracking, and asset pipelines, and adapted Hy4 for Unreal Engine, Unity, and Godot plus multiplayer development. The release acknowledges a gap between Hy4 and top proprietary models on internal benchmarks for research-level mathematics, long-horizon coding, model self-improvement, and advanced visual question answering, and says it will continue scaling pre-training and post-training.

RunAnywhere open-sourced an AWS Trainium2 inference engine for GLM-5.3-Flash, a 320B Mixture-of-Experts model with a 1M-token context, that serves the model on one trn2.48xlarge and rewrites each layer half as a single NKI kernel, or each whole layer for sparse-attention layers. On the same layout, a batch-8 decode step fell from 32.2 ms in the 17 October build to 24.2 ms on 1 November; the author attributes the gap to long chains of small vector-engine instructions at about 153 ns fixed cost each, HBM writes at every kernel boundary, and 125 cross-chip collectives per step. The remaining 24.2 ms includes a 6.1 ms memory floor, 10.0 ms of engines busy beyond it, 4.6 ms of waits inside kernels, and 2.8 ms of exposed collectives. An approximation remains in sparse-attention prefill on 1.2% of selections, shifting next-token log-probabilities by 0.012 to 0.021 nats on average and changing the top token at 3 of 448 positions; no task-level evaluation has been run on this build.

Libghostty 1.0 has been released as the first stable version of Ghostty&\x27;s embeddable terminal-emulation library. The release makes Ghostty&\x27;s terminal emulation available for reuse in other applications, targeting developers who would otherwise need to build that functionality themselves. The supplied material does not detail supported platforms, APIs, or compatibility.

Taobao’s MNN team has reportedly open-sourced MNN-Pocket-0.3B, an approximately 300-million-parameter model designed to run fully offline on phones. The release includes separate versions for pretraining, post-training, and on-device deployment, with an INT4 deployment file said to be 173 MB and a target peak memory below 300 MB. The team reportedly used a 28-layer structure, cut the vocabulary from 151K to 64K, and claims average decoding of about 350 token/s on an iPhone 18 Pro, against a target of more than 200 token/s. Those speed and memory figures are source claims from a brief secondary report and have not been independently verified.