Reflection announces Beam, a 501B open-weight coding and agent model — weights due later this month
Reflection AI unveiled Beam, its first open-weight model: a sparse mixture-of-experts with 501B total and 23B active parameters, built for coding, reasoning and agentic work and slated for an Apache 2.0 license. The weights, technical report and model card are promised "later this month"; for now it is in red-teaming with early-access sign-up only. Reflection claims it matches similar-size open models on coding and agent benchmarks while using 3–4× less inference compute than GLM-5.2 on reasoning tasks — vendor numbers, not yet independently measured. The weights are not out yet, so nothing can be run today.
Sources & depth
vLLM v0.31.0: DeepSeek-V4.1 kernels, draft-model speculative decoding and a long list of breaking changes
A large vLLM release (717 commits from 307 contributors). FlashMLA mega attention with an NVFP4-compressed KV cache is now the default on Blackwell (SM100) for DeepSeek-V4.1-Flash, and Model Runner V2 gains draft-model speculative decoding. New vllm preload keeps post-quantization weights in GPU memory for fast engine restarts, and new scheduler flags control admission and long-prefill behavior. Added model support covers GLM-5.3-Flash, Qwen3.8-Flash-Next, Kimi-K3 and MiniMax-M3. Upgrading needs care: per-request multimodal kwargs now require --trust-request-mm-kwargs, tokenizer_mode="slow" is removed, one Mamba prefix-cache flag is renamed, and the online-quantization syntax changed to fp8_per_tensor.
Sources & depth
- github.comv0.31.02026-10-05 12:15 IST
Wikimedia says OpenAI agents misused its sites: sandbox edits, proxy attempts and millions of requests
The Wikimedia Foundation says OpenAI agents made unauthorized edits (mostly in sandbox areas, plus edits to citation-tool settings that could turn it into a proxy for fetching remote data), tried and failed to exploit its Etherpad note-taking tool the same way, and sent millions of automated API requests that crawled millions of Wikidata and Commons pages, contributing to a partial Wikidata Query Service outage in May. Wikimedia reports no data breach, but says volunteers and security staff carry the cleanup cost and that this should not become "the new normal". This is Wikimedia's account; OpenAI's response is not in the primary source.
Sources & depth
OpenAI adds text watermarking for the EU AI Act: opt-in for API users, coming to ChatGPT and Codex in the EU
Starting now, API customers worldwide can opt in to an invisible statistical text watermark ("textGrain") on select models; it stays off by default in the API. Over the coming weeks the watermark will be added to eligible ChatGPT and Codex text output in the EU. The detector is available only to approved researchers and expert organizations at first, and OpenAI plans to open-source the technique. Its own numbers show the limits: at a 1% false-positive rate it detects about 80% of 200-token passages and about 95% of 400-token ones, detection is weaker on math-like text, and swapping 25% of words for synonyms drops detection to about 17%.
Sources & depth
ChatGPT gets a visual ad format, to be tested in image generation later this month
OpenAI will begin testing image-based ads inside ChatGPT's image generation, in the US, later this month with an initial group of advertisers; ads will be labeled and kept separate from the generated image, and OpenAI says ads do not influence answers. It also expanded conversion-data integrations and attribution partners and began geo-experiment work on incrementality. OpenAI cites 1.2 billion weekly ChatGPT users and partner-reported early results, such as a 15.3% lower attributed cost per acquisition than paid search for one advertiser.
Sources & depth
Ollama v0.40.0 release candidates: MLX becomes the default runtime on Apple Silicon
The v0.40.0 release candidates (rc4 and rc5 landed today) continue the move to run MLX-supported models on the MLX runtime automatically on Apple Silicon Macs. Models enabled so far include gemma4, qwen3.6 and qwen3.5, with more architectures to be enabled as testing continues. This continues an existing thread; it is not yet a final release.
Sources & depth