Reflection releases Beam, a 501B open-weight coding model under Apache 2.0
Reflection published Beam on 2026-10-16, its first model: a mixture-of-experts text model with 501B total and 23B active parameters, 256 experts per layer with 8 routed per token, and a 1M-token context window. It is released under the Apache 2.0 license, with BF16, FP8 and NVFP4 checkpoints, day-one support in Transformers, vLLM and SGLang, and hosted access through Baseten, Fireworks and Together; it is also listed on OpenRouter. Reflection reports 81.6% on SWE-bench Verified, 76.5% on SWE-Bench Pro v2-Hard, 80.1% on Terminal-Bench v2.1 and 90.6% on GPQA Diamond. These are the lab's own figures, not independent measurements.
Sources & depth
Claude recovers how the NIST elliptic-curve seeds were generated
An Anthropic internal research model, asked how the NIST curve seeds were made, worked it out in under half an hour without being given the known hash-of-a-phrase clue. The seeds for P-192, P-224, P-256, P-384 and P-521 turn out not to be hash outputs: they come from the C library's lrand48 generator seeded with a Unix timestamp, which Claude found by spotting a truncated linear congruential pattern and recovering generator states with lattice reduction. Six curves' base points trace to a random-number generator from FreeLIP. The origins of the B-163 seed and of the base points for the larger NIST curves remain unexplained. Anthropic states this is not a break or weakness in ECDSA curves; the $36,864 seeds bounty will be donated to the Internet Security Research Group.
Sources & depth
Anthropic details how Claude Code reviews its own code
With Claude writing about 90% of Claude Code's code, Anthropic says merged PRs more than doubled from June to September while the share reverted fell by more than 40%. Every PR gets CI, a parallel-agent AI review in which verifier agents vote on findings, and a security review; human approval then scales by a risk tier set by a classifier, from none for low-risk changes to two people for high-risk ones. Capping review rounds raised first-round passes from 65% to 78% and cut PRs needing five or more rounds from 8.9% to 1.5%. Anthropic acknowledges it lacks circuit breakers that tighten approval when reverts or incidents rise. All figures are Anthropic's own.
Sources & depth
NVIDIA RTX Spark laptops go on sale, from $2,599
NVIDIA's RTX Spark pairs a Blackwell RTX GPU with a Grace CPU (custom-designed with MediaTek) in one superchip, with up to 128GB of unified memory and up to 1 petaflop of FP4 AI performance, in slim Windows laptops. Systems from ASUS, Dell, HP, Lenovo, Microsoft and MSI are available globally from 2026-10-16, starting at $2,599. NVIDIA's developer blog pitches the machines for coding, testing and deploying local AI on a single device.
Sources & depth
Llama.cpp runs GGUF models on a jailbroken PS5
A community project, PS5LM, uses llama.cpp's Vulkan backend on a jailbroken PS5 (firmware 7.00 to 13.60) and serves the model over the home network through OpenAI- and Anthropic-compatible endpoints. About 11.4 GB of shared memory is available for model and context. The author reports about 55 tokens/s for Qwen 3.5 9B at Q4_K_M and about 20 tokens/s for Qwen 3.8 27B at 2-bit with 64k context; these are unverified, self-reported figures.
Sources & depth
Gemini 4 Argon is now on a StarCraft agent benchmark
StarSkirmish Bench, where language models play StarCraft: Brood War by writing code, has been refreshed with results for the flagship models of OpenAI, Anthropic, Google and Muse, including Gemini 4 Argon. It is a community benchmark, not a vendor result.
Sources & depth