AI releases
Reflection AI Beam Model: 501B Parameters, 23B Active
Reflection AI has announced Beam, an open-weight model with 501 billion total parameters, 23 billion active parameters and a planned Apache 2.0 release.
Prompt Blueprints / Topics
AI in the open.
Open-source and open-weight AI, local models and the tools built around them. Licences differ: open weights do not always mean open source.
Latest story
AI releases
Mistral Large 4: API Access and Open-Weights Timeline
Mistral announces Large 4, a natively multimodal model with 1 trillion total parameters, API access today and open weights planned for October.
Read the storyKeep reading
News
AI company news roundup: models, privacy and agents
A source-qualified roundup of reported changes involving Gemini, Kolibri-1, Claude, OpenAI, Mistral and Meta, including what remains unconfirmed.
News
Aleph Alpha Kolibri 1: Open Weights and 1M Context
Aleph Alpha’s Kolibri 1 is an Apache 2.0 open-weight MoE model focused on German and English, with long-context support and tool calling.
News
Xiaomi MiMo-V2.6 Agent Arena Rankings: Pro vs Flash
Arena reports different strengths for Xiaomi’s MiMo-V2.6-Pro and MiMo-V2.6-Flash, with Pro ranking higher and Flash costing less per task.
AI releases
GLM-5.3 UNCENSORED EXL3 3.0bpw: What the Release Includes
A source-based explanation of the community-hosted GLM-5.3 UNCENSORED EXL3 quant, its model lineage, reported tests, TabbyAPI setup and tool-calling caveat.
News
AI Product Updates Reported in the October 2 Daily Brief
A source-attributed roundup of reported updates involving Grok, Claude Code, Microsoft audio models, Perplexity, FLUX 3 Image, Cursor and OpenAI.
News
DeepSeek Harness Download: What the Preview Supports
DeepSeek Harness is an open-source preview with desktop, web, plugin, automation, and developer features. Here are the stated ways to access it.
News
GLM-5.3 Cybersecurity: Anthropic’s Test
Anthropic says GLM-5.3 can develop end-to-end exploits in controlled tests, while its open-weight release makes its safeguards easier to bypass or remove.
AI releases
Naive-N0.5-Flash: Inside NaiveAI’s 309B Open-Weight Model
Naive-N0.5-Flash is NaiveAI’s 309B open-weight MoE model for coding and AI R&D, with a native 1M-token context and the NaiveRT inference runtime.
Leaks & rumours
What the DeepSeek V5 Leak Actually Reports
An unverified X thread reports a possible DeepSeek V5 launch, rumored 2-trillion-parameter scale and Huawei Ascend training, but provides no official…
News
Qwen Intelligence: Alibaba’s Three Mobile Agents Explained
Alibaba describes three Qwen Intelligence agents for mobile planning, app use and creative generation. Here is what their reported benchmarks measure—and what…
AI releases
Xiaomi MiMo-V2.6 Pro and Flash: Reported Benchmarks
Xiaomi announces MiMo-V2.6 Pro and Flash with open model weights and reports mixed benchmark results against MiMo-V2.5 and named frontier models.
News
DeepSeek Reportedly Plans 2T and 8T Models on Huawei Chips
A post reports that DeepSeek is training a model described as 2T, planning an 8T model and expecting Huawei training chips in late 2026 or early 2027, but the…
AI releases
Kimi Code Desktop brings multiple coding agents into one workspace
Kimi describes Kimi Code Desktop as a workspace for managing multiple coding agents, running tasks in parallel and coordinating long-running programming work.
Opensource
ZCode Open Source Release: Interfaces and Security Status
Z.ai says ZCode is now open source and that reported security issues have been addressed. Its repository contains desktop, web and terminal interfaces, while…
AI releases
Qwen-Image-2.1: Open-Weight Image Generation and Editing
A grounded overview of Qwen-Image-2.1’s documented image-generation, editing, transparency, reference-image, and developer-integration features.
News
What Is Naive AI? Its Reported Open-Weight Model Explained
Naive AI is reportedly preparing its first open-weight language model, but its release timing, specifications and license remain unknown.