1. DeepReinforce Updates Ornith Model Family with 1.5 Release
Building on the previous Ornith-1.0 and 2.0 releases, DeepReinforce has launched the Ornith-1.5 model family. This update introduces a 9B dense model, a 35B MoE, and a 397B MoE, trained via an end-to-end self-improvement loop. The 1.5 release improves upon the performance of earlier versions, with the 397B model matching Claude Opus 4.8 benchmarks and the 9B model now supporting mobile deployment on iOS and Android.
- • The Ornith-1.5 family includes a 9B dense model, a 35B MoE, and a 397B MoE model.
- • The 397B MoE model matches Claude Opus 4.8 performance on key benchmarks.
- • The 9B dense model includes a mobile-optimized version for iOS and Android.
- • Models are trained using an end-to-end self-improvement loop for task generation and solution rollouts.
- • The family is available on Hugging Face in standard and GGUF formats.
This release provides developers with updated, more capable open-weights models for complex reasoning and agentic workflows, including new mobile-optimized options.
2. InclusionAI Expands Ling-3.0 Ecosystem with Base Model Checkpoints
Following the initial release of the Ling-3.0 model weights, InclusionAI has now open-sourced six base model checkpoints for Ling-3.0-tiny and Ling-3.0-flash. These checkpoints, which span pre-trained, mid-trained, and WSM-merged stages, provide developers with granular starting points for custom fine-tuning and continual pre-training. The models utilize Weighted Checkpoint Merging (WSM) to replace traditional learning rate decay, with the 124B parameter Flash variant demonstrating strong performance in coding and reasoning tasks.
- • The release includes six checkpoints covering pre-trained, mid-trained, and WSM-merged stages for Ling-3.0-tiny and Ling-3.0-flash.
- • Ling-3.0-tiny-base has 7.9B total parameters (1.3B active) and outperforms the older Ling-2.5-mini-base.
- • Ling-3.0-flash-base has 124B total parameters (5.1B active) and targets coding, reasoning, and long-context tasks.
- • The training process uses Weighted Checkpoint Merging (WSM) instead of traditional learning rate decay.
- • Fully post-trained versions remain available as Ling-3.0-tiny and Ling-3.0-flash.
These intermediate checkpoints offer developers greater flexibility for fine-tuning and continual pre-training compared to the previously released final weights.
3. TrueFoundry Open-Sources TrueForge AI Agent Harness
TrueFoundry has open-sourced TrueForge, an MIT-licensed AI agent harness designed to help enterprise developers build and manage cost-efficient, vendor-neutral agents. TrueForge employs context engineering techniques—such as offloading large tool results to files and compacting conversation histories—to minimize API costs. Additionally, its sandbox-as-a-tool architecture provisions isolated environments on demand for secure code execution, offering a highly efficient alternative to managed agent platforms.
- • TrueForge is released under the MIT License by TrueFoundry.
- • It utilizes context engineering (offloading large tool results to files, compacting history) to reduce model spending.
- • The harness features a sandbox-as-a-tool architecture that provisions isolated environments only when needed.
- • TrueFoundry claims TrueForge with GLM-5.2 is 75% cheaper than Claude Managed Agents with Claude Opus 4.8 on Enterprise-Bench.
- • It can be paired with TrueFoundry's commercial AI Gateway for access controls and governance.
It gives developers a vendor-neutral, cost-efficient framework to deploy and manage AI agents with built-in sandboxing and context optimization.
4. Miles v0.1 Released for Production-Level Agent Post-Training
Miles v0.1 has been released as an open system for production-level post-training of AI agents. The framework enables developers to run multiple agent instances in isolated sandbox environments, automatically scoring their performance and feeding the results back into a reinforcement learning loop. Miles supports asynchronous training and hot-swapping model updates to active workers, allowing continuous agent improvement without pipeline downtime.
- • Miles v0.1 is an open system for post-training AI agents using reinforcement learning.
- • It runs multiple agent copies in isolated environments to score task performance and feed results back into training.
- • The system allows distributing updated models to workers without stopping the training pipeline.
- • It includes components for rollout, sandboxing, asynchronous training, replay, model-update, and multi-hardware scaling.
It provides developers with an automated pipeline to continuously train, evaluate, and deploy updated agent models without stopping production workflows.
5. CHAP 0.2 Released for Auditable Human-Agent Collaboration
Brightbeam AI has released CHAP 0.2 (Collaborative Human-Agent Protocol), an open protocol designed for structured, auditable collaboration between humans and AI agents. CHAP stores human edits to agent drafts in queryable, content-hashed envelopes to maintain a single source of truth. The release includes reference implementations in Python and TypeScript, support for MCP, and framework bridges for popular libraries like LangGraph, Pydantic AI, and LlamaIndex.
- • CHAP (Collaborative Human-Agent Protocol) 0.2 is available as a public draft.
- • The protocol stores human edits to agent drafts in queryable, verifiable envelopes linked by content hash.
- • It includes reference implementations in TypeScript and Python, a conformance harness, and support for MCP and A2A transports.
- • It provides framework bridges for LangGraph, Pydantic AI, AG2, LlamaIndex, and Google ADK.
- • Optional security features include OIDC-bound signatures and external transparency-log anchoring.
It provides developers with a standardized, secure framework to build human-in-the-loop agent workflows with built-in audit trails.
6. Coding Tools Adopt AGENTS.md for Codebase Comprehension
AI coding tools including Cursor, Codex, and Amp are adopting a new standardized Markdown format called AGENTS.md to improve codebase comprehension for AI agents. Designed as a tool-agnostic alternative to the Claude Code-specific CLAUDE.md format, AGENTS.md provides a unified way for developers to document codebase structure, guidelines, and context specifically for autonomous coding agents.
- • Codex, Amp, and Cursor are adopting AGENTS.md as a standardized Markdown file format.
- • The format is designed to provide a unified way for coding agents to understand a codebase.
- • It serves as a tool-agnostic alternative to CLAUDE.md, which is specific to Claude Code.
It gives developers a tool-agnostic way to document their codebases specifically for AI coding agents, replacing proprietary formats.
7. Modular Open-Sources Mojo 1.0 and Expands Platform Support
Modular has announced that the Mojo 1.0 programming language is now fully open source under the Apache 2.0 license. Alongside the open-source release, Modular launched Modular Cloud for shared and dedicated deployments and expanded its platform's hardware support to include AWS Trainium, Google TPUs, and Qualcomm accelerators. Additionally, the MAX license has been updated to remove device usage restrictions, transitioning to a source-available model under a new open alliance program.
- • Mojo 1.0 is now fully open source under an Apache 2.0 license.
- • Modular Cloud is publicly available at console.modular.com, offering shared endpoints and dedicated deployments.
- • The Modular Platform expanded hardware support to AWS Trainium, Google TPUs, and Qualcomm Cloud AI 100 and Dragonfly accelerators.
- • Modular is collaborating with Microsoft to develop native Windows support for Mojo.
- • The MAX license was updated to remove device usage restrictions and will be source-available.
It gives developers a fully open-source, high-performance language for AI systems engineering with broader hardware compatibility.
8. Leveraging PostgreSQL for Vector Search and AI Workflows
PostgreSQL continues to expand its utility in AI engineering through specialized extensions. By leveraging the pgvector extension for vector embeddings and the pgai extension for direct LLM indexing and model calls, developers can consolidate their database, vector search, and LLM integration layers. This ecosystem, combined with native support for JSON storage, full-text search, and high-performance caching via unlogged tables, allows teams to build robust AI backends entirely within a single database.
- • The pgvector extension enables PostgreSQL to serve as a vector database for AI workflows.
- • The pgai extension facilitates indexing data and calling LLM models directly within PostgreSQL.
- • PostgreSQL supports full-text search, JSON document storage, partitioning, and common table expressions.
- • It can function as a high-performance queueing system using SELECT .. FOR UPDATE and SKIP LOCKED.
- • Unlogged tables can be used for non-persistent high-performance caching.
It helps developers simplify their stack by using a single, reliable database for relational data, vector embeddings, and LLM integration.
9. Developers Release Tools to Bypass Claude's New Watermarks
Building on the recent implementation of Google's SynthID-Text watermarking in Claude models, developers have already created methods to circumvent the technology. Within four hours of the rollout, developer Guillaume Meyer released an open-source override tool designed to strip the watermarks. Other community-identified bypass methods include paraphrasing, translating, or reordering text, highlighting the ongoing tension between AI provenance efforts and user-side modification.
- • Anthropic recently implemented SynthID-Text watermarks in Claude models to comply with the EU AI Act.
- • Developer Guillaume Meyer released an open-source tool to strip these watermarks within four hours of the announcement.
- • Community members have identified additional bypass techniques, including text reordering and translation.
- • Anthropic maintains that the watermarks do not affect response quality and plans to release a detection API.
It demonstrates the immediate community response to Anthropic's new watermarking implementation and provides developers with tools to remove synthetic identifiers from model outputs.
10. Inco AI Launches DFlash 2, Building on Original Speculative Decoding Research
Building on the original DFlash speculative decoding research, Inco AI has released DFlash 2. This update introduces a lightweight path selector and a local convolution module to mitigate suffix decay, adding 16.5 million parameters while boosting throughput by 16% to 25% over the original DFlash. The technology is now available for integration, with support already added to llama.cpp and vLLM, enabling Qwen3.8-27B to achieve up to 3.4x the throughput of standard autoregressive decoding.
- • DFlash 2 improves throughput by 16% to 25% over the original DFlash.
- • New features include a lightweight path selector and a 16.5M parameter local convolution module to address suffix decay.
- • Inco AI released DFlash 2 drafters for Qwen3.8-27B and Meta's Muse Glimmer on Hugging Face.
- • Llama.cpp pull request #27342 adds support for DFlash 2, showing up to a 3x speedup for Qwen3.8-27B on an RTX 6000.
- • A vLLM-based implementation achieved 138 tokens per second for single-user requests on an RTX 3090.
It provides a significant performance upgrade for developers already utilizing the DFlash speculative decoding framework for local LLM inference.
11. Unsloth Releases Dynamic v3.0 GGUF Quantizations
Unsloth has introduced its Dynamic v3.0 quantization format, releasing updated GGUF models for Qwen3.8-27B. The new quantization method leverages a refined imatrix calibration dataset to optimize for agentic coding, chat, and multilingual performance. By avoiding QAT and QAD, Unsloth delivers a 10% accuracy improvement over previous versions, including an ultra-low 1-bit quantization that retains over 70% accuracy on hardware with only 8GB of RAM.
- • Dynamic v3.0 GGUFs are compatible with llama.cpp and Unsloth Desktop.
- • The quantization method uses a high-quality imatrix calibration dataset refined for agentic coding, chat, and multilingual tasks.
- • The process relies entirely on post-training quantization without quantization-aware training (QAT) or quantization-aware distillation (QAD).
- • A new 1-bit UD-IQ1_S quant is 6.2GB in size, retains 72% to 77% accuracy, and can run on 8GB of RAM.
- • Unsloth removed the MTP module from smaller quants under UD-Q2_K_XL to save approximately 500MB of disk space.
It enables developers to run highly accurate, quantized models locally on consumer hardware with minimal memory footprints.
12. FreeToken Enables Efficient Local MoE Serving
Researchers have introduced FreeToken, a technology designed for efficient, edge-native serving of Mixture-of-Experts (MoE) models. By dynamically remapping experts, model state, CPU/GPU workloads, and agent state reuse to match the host machine's available memory and bandwidth, FreeToken enables consumer hardware to run massive models. The system supports over 20 MoE models, ranging from 35B models on an 8GB laptop GPU up to a 753B GLM model on a single workstation GPU.
- • FreeToken dynamically remaps experts, model state, CPU/GPU work, and agent state reuse based on available memory.
- • The technology supports more than 20 MoE models.
- • Supported setups range from running 35B models on an 8GB laptop GPU to a 753B GLM model on a single workstation GPU.
It allows developers to run massive MoE models locally on standard workstations or laptops by optimizing memory and bandwidth.
13. Software Engine Enables Native FP4/FP8 on Tesla V100 GPUs
A developer has released 'v100-skinny', a software engine designed to run NVFP4 and FP8 model weights natively on Volta-generation Tesla V100 GPUs. By utilizing custom Quantized Processing Network (QPN) kernels, the engine translates low-precision weights into FP16 formats compatible with Volta Tensor Cores, bypassing the hardware's lack of native FP4/FP8 support. In testing with Qwen 3.8, a cluster of four V100s achieved 219.1 tokens per second, matching the single-request decode performance of a modern RTX 5090.
- • The 'v100-skinny' engine enables four Tesla V100 GPUs to achieve a decode throughput of 219.1 tokens per second on Qwen 3.8.
- • The performance matches the single-request decode speed of an RTX 5090 running the NInfer engine.
- • The system uses Quantized Processing Network (QPN) kernels to translate FP4 and FP8 weights into FP16 formats compatible with Volta Tensor Cores.
- • Version 1.1 provides specific SM70 execution paths to keep the model's mixed FP4/FP8 allocation intact.
- • The setup was built using four V100 GPUs acquired for approximately A$600 in hardware costs.
It allows developers to repurpose cheap, older enterprise GPUs to match the single-request decode performance of modern flagship cards.
14. OpenAI Tests Private Safety Processing for Zero Data Retention
OpenAI has begun testing Private Safety Processing with select enterprise customers. This new technique is designed to detect misuse patterns and safety violations while fully preserving zero data retention (ZDR) protections, allowing organizations with strict compliance needs to leverage OpenAI's frontier models safely.
- • The technique is called Private Safety Processing.
- • It is designed to identify misuse patterns while preserving zero data retention protections.
- • The feature is currently being tested with early enterprise customers.
- • It allows businesses to use OpenAI's most advanced models without requiring OpenAI to retain their data.
It allows developers in highly regulated industries to use advanced OpenAI models without compromising strict data privacy and retention requirements.