DeAI Daily Brief — 5 November 2026

Today in DeAI: Together AI claims an agent-driven compiler lifts open-model serving throughput up to 2.6x, Tencent announces the Hy4 model while withholding its weights, and Akash Network governance approves a Cosmos-to-Solana migration.

Together claims its agent compiler rewrites serving code fast enough to matter

Together AI introduced Together Compiler, a loop where AI agents generate and test serving optimizations against "persistent engines" that hot-reload candidate code in seconds. The company claims up to 2.6x higher tokens-per-second over the best open engine configurations on Kimi K3, GLM 5.3, MiniMax M3.1-Flash, and DeepSeek v4.1-Flash, plus claimed InferenceX low-latency records of 1111 and 1205 TPS. Every figure is a vendor self-report; the SemiAnalysis leaderboard is where it gets tested. Why it matters: if agent-driven recompilation is real in production, it is one of the mechanisms behind falling serving costs — and a new benchmark gap between managed providers and self-hosters (our full coverage).

Tencent's Hy4 is announced, but the weights are only "planned"

Tencent's Hy4 research page describes a 770B-parameter mixture-of-experts model with over 1M-token context, first-time vision and native computer use, live on Tencent Cloud and OpenRouter. The page says the team "plans to open-source Hy4's model weights"; no weights repo exists as of November 4. Why it matters: this is the announcement-without-weights pattern that made the Kimi K3.1 Pro deadline a story — self-hosters cannot touch Hy4 today, and the license terms are unknown.

Akash votes to leave Cosmos for Solana

Akash Network's Proposal 346 closed November 2 with a reported 97.27% in favor of migrating the decentralized compute marketplace to Solana within 24 months, swapping AKT one-to-one into a native Solana token. Why it matters: a live GPU marketplace is moving its settlement and identity layer mid-operations; the swap mechanics and provider staking paths are what to watch. Our decentralized inference networks overview will track what actually breaks.

Anthropic's FlashProve makes verifiable inference 70-95x more expensive

Anthropic's FlashProve release (Apache 2.0, published November 2) demonstrates zkSNARK proofs that a response came from committed model weights, at roughly 110 tokens/second proof generation on one H100 for a 400B model. The cost: about 70-95x the GPU time of plain vLLM inference, on a research prototype capped near 1,024 tokens with no deployment planned. Why it matters: the bait-and-switch trust problem is real for API buyers, and this is the first credible cryptographic answer from a major lab — but it is a research signal, not deployable tooling yet.

KeyHunter reports 2,233 LLM API keys exposed, LiteLLM risk in the same window

SOCRadar's KeyHunter operation reports 2,233 LLM API keys harvested from exposed environment files, GitHub repos, and open gateways, with 223 reportedly usable across major providers — a threat-intel vendor's own research count, not an independent audit. It lands in the same window as the LiteLLM v1.107.0 supply-chain compromise we covered and a Railway env-var access warning. Why it matters: if you self-host AI stacks or stuff provider keys into flat env files, rotate, scope, and patch past v1.107.0 now.

Watching tomorrow

Whether Tencent ships the Hy4 weights repo it promised, and whether any third party reproduces Together's InferenceX numbers.