
Reflection's Beam, Cloudflare's Clef, and the new wave of open-weight models
Reflection released Beam on 5 October 2026, a 501-billion parameter open-weight model built to challenge Chinese competitors in coding and agentic tasks.

Reflection released Beam on 5 October 2026, a 501-billion parameter open-weight model built to challenge Chinese competitors in coding and agentic tasks.

Reflection AI unveiled Beam, its first frontier open-weight model, on Monday, October 5. The startup claims the 501-billion-parameter system matches leading Chinese open models on advanced reasoning benchmarks while using 3-4x less inference compute.

On October 5, 2026, Reflection AI released Beam, a 501-billion-parameter open-weight model. The company says it matches top Chinese rivals on reasoning benchmarks while using much less inference compute.

Two coding agent projects, Reika and Jeff-Code, both published on 5 October and both claim the same thing: the bottleneck for small local models is the harness around them, not the model. Liquid AI's d1, out the same day, takes a different route and replaces the language model with a decision model.

Two coding agent releases on 5 October make the same point: for small local models, the scaffolding around the model is the bottleneck, not the model. Both projects publish concrete numbers.
Reflection released Beam on 5 October, a 501B-parameter open-weight model. Its own benchmark table shows it trailing Kimi K3 by 16.6 points on agentic terminal coding. A UC Riverside study published on 23 September says confidence and correctness sit in different parts of a model, so systems can be fluent and wrong at once.
Reka released a research preview of Rho-1 on 5 October, according to the company's lab post. The 19B model handles text, images, video and robot actions inside one network. Liquid AI, Interfaze and others offered smaller, cheaper multimodal bets the same day.
IronBee AI says its Express browser agent bought an iPhone 15 Pro in a demo shop in 6.7 seconds for $0.00054. On a second run it caught a checkout that claimed success while the order behind it had failed. The company published the engineering write-up on 5 October.
On 5 October, Liquid AI launched d1, a "decision model" that returns probabilities instead of generated text. The company says it answers in 200 to 300 ms.
A benchmark revised on arXiv on 1 October found that memory agents handle structured recall tasks when prompted how to organise their memory. Modern LLMs do not always recognise the structure on their own.
Liquid AI released d1, its first decision model, on 5 October. It handles text and images, and Liquid says it costs 19x to 200x less than GPT-6.1 Sol and Claude Opus 5.5 and answers faster on every task it tested. The security angle is real but indirect: d1 is not an agent that defends networks, it is a classifier that could replace costly model calls inside one.
Liquid AI released d1 on 5 October, its first decision model with vision support. The company claims it matches or beats GPT-6.1 Sol on four of six real applications while costing 19x to 200x less. The same day brought a 501B open-weight model from Reflection, a 0.8B coding router, and a 64-bit Expat fix. All of it points at cheaper, faster inference rather than bigger models.
Two benchmark releases in the past 24 hours put hard numbers on what AI agents actually do. GitHub's ReviewBench covers AI code review, and Veris's VAmoS Pro board covers voice agents. Both show the same gap: the scores are narrow, and the vendors know it.
Amazon said on 5 October it will put more than $1 billion into workforce education, energy and water programmes in the communities that host its data centres. The pledge answers mounting local opposition as several US states prepare data centre ballot measures.
A cluster of new developer tools for AI coding agents landed this week. Threadnote published a five-task study claiming 65.62% fewer lifecycle tokens per verified completion when an agent resumes from a saved handoff instead of raw files.
Liquid AI released d1 on 5 October, a decision model that now reads images as well as text and, the company says, answers a text question in 200 to 300 milliseconds without generating a single token.
India's CERT-In has been given access to Mythos, an unreleased Anthropic model, for cyber defence testing, according to reporting on the arrangement.
On 5 October, GitHub published ReviewBench, an open code review benchmark modelled on an analysis of 103.9 million pull requests and containing 219 cases. In September, UC Riverside researchers described a way to separate a model's confidence from its correctness.