DeAI Daily Brief — 6 Oct 2026

Today in DeAI: Reflection AI launches Beam with weights promised this month, OpenAI ships EU-mandatory text watermarking with published detection limits, Iterate.ai launches attestation-gated confidential inference, and Together ships a free CLI that points coding agents at open-weight models.

Reflection AI launches Beam: 501B open-weight MoE, weights promised this month

Reflection AI — the Nvidia-backed startup founded by ex-DeepMind researchers Misha Laskin and Ioannis Antonoglou — introduced Beam on October 5: a sparse mixture-of-experts model with 501 billion total parameters and 23 billion active, pretrained on 23.8 trillion tokens and refined with a reinforcement-learning run that used 10,500 Nvidia GB300 GPUs over four weeks and generated over 100 million rollouts. The company says Beam matches Z.ai's GLM-5.2 on advanced reasoning while using 3-4x less inference compute, an estimate its own methodology note says excludes prefill and serving overhead, and its self-reported table puts Beam at 80.1 on Terminal-Bench v2.1 versus 88.3 for Kimi K3 and 90.6 for DeepSeek V4.1 Flash. Apache 2.0 weights, a technical report, and FP8/NVFP4 quantizations are promised later this month; today the only access is an early-access waitlist, and no independent benchmark of Beam exists. Why it matters: the day-one checkpoint supply for frontier-capable open weights has run through Kimi, GLM, DeepSeek, and Qwen, and a US entrant widens it only when the weights actually ship. (Sources: Reflection AI, Reuters via The News Tribune) — our full coverage

OpenAI ships textGrain: EU-mandatory text watermarking with published detection limits

OpenAI published its approach to the EU AI Act's text-provenance requirement on October 5: an invisible statistical watermark called textGrain rolls out to eligible ChatGPT and Codex output in the EU over coming weeks, API customers globally can opt in starting today with watermarking off by default, and detector access goes only to approved researchers and expert organizations. The post's own numbers define the limits: at a 1% false-positive target, the detector finds watermarks in about 80% of 200-token passages and about 95% of 400-token passages, with substantially lower rates for low-entropy content like math, and replacing 25% of words drops detection from 92% to 17%. An open-source release is planned but not shipped, so third parties cannot yet verify a detection result independently. Why it matters: EU-facing closed-API output now carries a machine-readable provenance signal, while open-weight output that EU users run themselves structurally cannot — a compliance asymmetry between hosted APIs and self-hosted weights. (Source: OpenAI) — full coverage tomorrow

Iterate.ai launches Lifeboat: attestation-gated confidential inference at $499.99/month

Iterate.ai launched Lifeboat, an LLM inference engine with confidential computing built in: in its Confidential Computing edition, no request is served until hardware attestation passes, covering AMD SEV, Intel TDX, and Nvidia's confidential computing mode on H100, B200, and GB300 GPUs, with model weights sealed in the trusted execution environment and encrypted in use. The company claims 2-6x more concurrent AI agent sessions per GPU; its own test held 2,048 concurrent sessions on a single Nvidia RTX PRO 6000 Blackwell running a Qwen 30B-A3B model, and a memory-pressure run kept 99th-percentile time to first token at 1.5 seconds versus a baseline of 189 seconds. All throughput figures are vendor-run, and SiliconANGLE had no independent test. Pricing is $49.99/month Standard and $499.99/month Confidential Computing, each with a seven-day free trial. Why it matters: attestation-gated inference is becoming a productized feature rather than a bespoke deployment, joining Cohere's Model Vault and NEAR AI in the TEE-inference pattern. (Source: SiliconANGLE) — full coverage tomorrow

Together ships Together Link: a free CLI that runs open models inside Claude Code, Codex, and OpenCode

Together AI shipped Together Link, a beta CLI that connects existing coding agents and desktop apps to models hosted on Together, with explicit tier mapping in Claude Code sessions: Opus routes to Moonshot AI's Kimi K3, Fable to Z.ai's GLM 5.3, and Sonnet to DeepSeek V4.1 Flash, with per-model overrides. A default auto router picks a Together model per request and, when an Anthropic API key is present, sends the hardest requests to Claude Opus billed to that account. The CLI is free; the hosted models bill through Together under Together's data-handling terms, which are not the labs'. Why it matters: the harness stays, the weights source swaps, which is a practical cut in the switching cost of running open models in agent workflows. (Source: Together AI docs) — full coverage tomorrow

Watching tomorrow

Beam's Hugging Face org page: Reflection has committed to Apache 2.0 weights, a technical report, and FP8/NVFP4 quantizations later this month, and independent scoring that follows will either support or break the 3-4x efficiency claim.