Inside: Zyphra converts autoregressive to diffusion → Nous cuts pre-training 2.5x → Cline open-sources its agent engine → Supertonic ships 31-language on-device TTS

1. Zyphra Releases ZAYA1-8B-Diffusion-Preview: First MoE Diffusion Model Converted From an Autoregressive LLM — Up to 7.7x Speedup — Zyphra has released ZAYA1-8B-Diffusion-Preview, the first diffusion language model converted directly from an autoregressive LLM — and the first diffusion LLM trained on AMD hardware. Instead of generating one token at a time, the model diffuses blocks of 16 tokens simultaneously. That delivers a 4.6x speedup with a lossless sampler and 7.7x with an aggressive sampler — on a 760M active / 8.3B total parameter MoE architecture. No quality degradation. No retraining from scratch.

2. Nous Research Releases Token Superposition Training — 2.5x Faster LLM Pre-Training, Zero Architecture Changes — Nous Research has released Token Superposition Training (TST), a modification to the standard LLM pretraining loop that delivers a 2–3x wall-clock speedup at matched FLOPs — with no changes to the final model, optimizer, tokenizer, or architecture. Validated across 270M to 10B parameter models. You get the same model at the end, trained in half the time.

3. Mistral Vibe now moves coding agents to the cloud so you can run several in parallel and stop being the bottleneck on every step the agent takes. Each session runs in an isolated sandbox. Start from the Vibe CLI or Le Chat, inspect file diffs, tool calls, and progress states as they run, and come back to a finished branch or draft PR. Already working locally? Teleport your session to the cloud and keep going without losing context. Available on Le Chat Pro and Team. Get Started with Vibe _(promoted)

4. Cline Open-Sources Its Core Agent Runtime — The Same Engine Powering VS Code, JetBrains, CLI, and Kanban — Cline has rebuilt its core agent harness from scratch and released it as the Cline SDK — an open-source agent runtime now powering Cline's CLI, Kanban, VS Code extension, and JetBrains plugin. Any developer can now build on the same engine running one of the most widely used AI coding agents in production. Custom tools, multi-agent orchestration, and full programmable control — available via the SDK today.

5. Poetiq's Meta-System Hits SOTA on LiveCodeBench Pro — No Fine-Tuning, No Special Model Access — Poetiq's Meta-System automatically builds a model-agnostic coding harness that improved every LLM tested on LiveCodeBench Pro — reaching a new state-of-the-art without fine-tuning, retraining, or privileged model access. The system builds its own harness recursively, then applies it across models. Works on GPT-5.5, Claude, Gemini — any model you plug in gets better. This is recursive self-improvement in practice.

6. Supertone Releases Supertonic v3 — On-Device TTS, 31 Languages, No Cloud Required — Supertone has released Supertonic v3, an on-device text-to-speech model that runs entirely via ONNX Runtime — no cloud, no API call, no latency penalty. Expands from 5 to 31 languages, adds expression tags for emotion control, and significantly reduces repeat and skip reading failures. Faster, more stable, more expressive — and fully local. Open-weight ONNX assets on Hugging Face now.

Best AI Coding Agents in 2026 — Benchmark-Driven Rankings — MarkTechPost published a comprehensive benchmark-driven ranking of AI coding agents. Claude Code on Opus 4.7 leads SWE-bench Verified at 87.6%. GPT-5.5 tops Terminal-Bench 2.0 at 82.7%. Full breakdown covers Cursor, Devin, Aider, Cline, OpenHands — with verified scores, pricing, and architecture tradeoffs. The definitive 2026 reference for picking a coding agent.

Fastino Labs Open-Sources GLiGuard: 300M Parameters, Beats Models 90x Its Size — GLiGuard scores 87.7 average F1 across nine safety benchmarks — within 1.7 points of the best model that is 90x its size — while running up to 16x faster. Covers prompt safety, response safety, 14-category harm classification, and jailbreak detection. Apache 2.0. Production-ready today.