# Reflection's Beam, NIST curve seeds traced, and RTX Spark > Four primary pages I read today: Reflection's 501B open Beam model, Anthropic on ECDSA seeds, Google's wet-lab agent papers and NVIDIA's RTX Spark. Published 16 Oct 2026. Coverage: 15 Oct 2026 to 16 Oct 2026. By Ali Asif. ## Items 1. **Welcome Beam: a 501B open model you can own, customize, and scale** (Reflection, Open weights): Apache 2.0 sparse mixture-of-experts with 501B total and 23B active parameters, a 1M token context, and BF16, FP8 and NVFP4 checkpoints with day-0 support in transformers, vLLM and SGLang. [Primary source](https://huggingface.co/blog/reflection-beam) 2. **Claude reveals ECDSA curve origins** (Anthropic, Frontier): An internal research model traced the NIST curve seeds to lrand48 outputs seeded with Unix timestamps, with CC0 code to reproduce it and the authors stating it is not a cryptographic break. [Primary source](https://www.anthropic.com/research/ecdsa-curve-origins) 3. **Grounding AI agents in biological reality: Building blocks for autonomous wet labs** (Google Research, Science): Three papers on a digital twin for liquid-handling robots, closed-loop aptamer design and faster protein binder design, with human review of every protocol. [Primary source](https://research.google/blog/grounding-ai-agents-in-biological-reality-building-blocks-for-autonomous-wet-labs) 4. **Code, Test, and Deploy on One Machine with NVIDIA RTX Spark** (NVIDIA, Agents and coding): A GB10-based local AI development platform with up to 128GB unified memory, Windows on Arm software and the open Nemotron 3.5 Bolt 74B-A9B model quantized to 24GB. [Primary source](https://developer.nvidia.com/blog/code-test-and-deploy-on-one-machine-with-nvidia-rtx-spark) Four pages today, all primary, and none of them a rerun of this week's earlier releases. Numbers below are the publishers' own. I haven't reproduced any of them. ## Beam, a 501B open model from Reflection Reflection released Beam-1.0-501B-A23B under Apache 2.0. It is a sparse mixture-of-experts model: 501B parameters in total, 23B active per token, 52 decoder layers, 256 experts per MoE layer with 8 routed per token. Context is 1M tokens, text in and text out. Sliding-window attention covers 512 tokens in 43 layers, and the other 9 use full context. What caught my eye is the checkpoint spread. BF16 is about 1 TB in safetensors, FP8 targets Hopper and Ampere, and NVFP4 (about 280 GB) targets Blackwell. The team reports 81.6% on SWE-bench Verified in BF16, 80.2% in FP8 and 80.7% in NVFP4. Terminal Bench v2.1 reads 80.1%, 78.9% and 78.2%. AIME 2026 is 97.7%, 98.0% and 97.5%, and GPQA Diamond 90.6%, 90.3% and 89.3%. If those hold up independently, quantizing costs very little here. Day-0 support covers transformers (5.20.1 or later), vLLM and SGLang, plus Hugging Face Inference Providers (Baseten, Fireworks, Together). ## Anthropic traces the NIST curve seeds NIST adopted ECDSA in 2000, and the seeds for its curves came from the NSA with no explanation. Anthropic's Frontier Red Team says an internal research model worked out their origin in about half an hour. The seeds are not hash outputs. They come from the C library's lrand48 generator seeded with a Unix timestamp, and the prime-curve seeds map to timestamps between May 1997 and April 1998. The authors call this neither a break nor a weakness, and say it makes a backdoor less plausible. They also list what stays open: the base points for P-224, P-256, P-384, P-521 and the binary curves, and the B-163 seed. The $36,864 bounty goes to the Internet Security Research Group, and the Python to reproduce the seeds is released under CC0. ## Google's building blocks for wet labs Three papers. VERID is a digital twin of a liquid-handling workcell on PyLabRobot, with virtual dry runs and checks that block invalid commands. In a self-play campaign to reconstruct blinded dye mixtures, accuracy improved more than fivefold. In the aptamer work, α-thrombin hit rates went from 0% in round one to 26.0% and then 42.1%, at 98.5% specificity. The agents also caught their own mistake on myoglobin, where specificity fell to 19.1% before a revised rule set brought it back to 72.4%. Epitope excision sped up each binder-design cycle 8.6 times, and top designs reached single-digit nanomolar binding to FAP. Google says humans authorize every sequence, protocol and robot run. ## NVIDIA's RTX Spark RTX Spark is NVIDIA's single-machine platform for local agent development, built on GB10 with up to 128GB of unified memory and 1 petaflop of FP4 compute. NVIDIA's comparison against an Intel Core Ultra X9 388H shows 4.6x on a local coding-agent benchmark. The release also includes Nemotron 3.5 Bolt 74B-A9B, quantized to 24GB. Laptops are available now, and small desktops arrive in November. ## What I'd watch next Whether anyone outside Reflection reproduces the Beam scores at NVFP4, and whether the open ECDSA questions get answered. For the wet-lab papers, I want to see the digital-twin idea used on someone else's hardware. ## My take Beam is the one I would try first, because the same weights come in three precisions and the reported scores barely move between them. The ECDSA post is the most careful of the four. The authors say plainly what is still unexplained. I have read all four pages and run none of it. ## Sources - https://huggingface.co/blog/reflection-beam - https://www.anthropic.com/research/ecdsa-curve-origins - https://research.google/blog/grounding-ai-agents-in-biological-reality-building-blocks-for-autonomous-wet-labs - https://developer.nvidia.com/blog/code-test-and-deploy-on-one-machine-with-nvidia-rtx-spark Canonical: https://heyaliasif.com/ai/2026-10-16-reflection-beam-ecdsa-seeds-wet-lab-rtx-spark