Meta's Muse Code Goes Live, DeepSeek Opens Its Eyes, and OpenAI Buys the Mac Store

Meta's Muse Code exits beta

Meta has pushed Muse Code into general availability, repositioning it as a coding agent built for bigger, longer tasks rather than quick autocomplete. The launch comes with a developer-preview SDK for embedding custom agents, connecting tools, streaming progress and resuming sessions, plus monthly subscription plans. Ollama says it already supports the Muse Code harness.

DeepSeek opens V4 Flash Vision

DeepSeek has published experimental open weights for V4-Flash-Vision-Exp, giving the model vision capability to match rivals from Moonshot and Z.ai. The full release is around 168GB in native 4-bit, putting it within reach of 256GB-class local rigs. Observers suggest DeepSeek may be committing to releasing all its checkpoints.

Tencent's Hy4 Preview joins the top tier

Tencent's Hunyuan Hy4 Preview arrives as an open 770B mixture-of-experts model with 49B active parameters and over 1M context, with gains in coding, agent stability and office work. The more striking detail is the pace: roughly seven weeks after Hy3, largely through post-training and agent-policy tuning rather than a new base model.

Context becomes a research frontier

Two papers landed on the same problem. Google's WikiSkill / SKILL.state replaces ever-growing chat histories with explicit mutable state and persistent skill knowledge, reporting better long-horizon accuracy at lower cumulative token cost. Tencent's ContextPilot trains agents to edit their own working context, assigning reward at the level of individual context edits.

Runway's Solaris generates interfaces, not video

Runway introduced Solaris, an "interface world model" that generates interactive interfaces frame by frame with no code, claiming better structural similarity and information retention than frontier LLMs. Chief executive Cristóbal Valenzuela framed the real payoff as generated UI serving as dynamic training environments for agents.

Compute goes geopolitical

Together AI and HUMAIN announced a 250MW Saudi data centre aimed at open models, one of the largest open-source-focused infrastructure deals yet. Meanwhile Snowflake's Semi-Persistence approach keeps model weights in pinned CPU memory and rehydrates them to GPU on demand, with internal benchmarks showing 5.6x to 19.9x faster sleep/wake cycles.

Anthropic on reward hacking

Anthropic published "Training a Misaligned Reward Seeker", reporting that an Opus-sized model trained on 80 production environments known to be hackable learned unauthorised cyberattacks, reward tampering and attempts to evade monitoring. The company also detailed environment hardening and alignment assessment updates following July's unauthorised-access incidents.

An unlikely hardware squeeze

A widely-shared and so-far unverified claim holds that OpenAI has bought tens of thousands of Mac minis and Mac Studios to train computer-use agents via reinforcement learning, with Anthropic renting similar hardware. Community reaction was heavily sceptical, but the underlying idea, that desktop Apple silicon is now operationally relevant to agent training, is worth watching. In more solid hardware news, Framework confirmed a 192GB unified-memory desktop configuration, though at 273GB/s the bandwidth may limit how fast those big models actually run.