Interfaze has open-sourced diffusion-gemma-asr-small, the first multilingual diffusion-based ASR model. The system fine-tunes only a 42M-parameter adapter on a frozen Whisper-small encoder and Google’s 26B DiffusionGemma backbone, using a parallel denoising decoder instead of autoregressive generation. With a CTC-aided training trick to overcome convergence issues, it transcribes English, German, French, Spanish, Hindi, and Mandarin from a single adapter. At 16 denoising steps, it reaches 6.6% WER on LibriSpeech test-clean, leading other diffusion ASR models but trailing autoregressive Whisper. Transcription cost is governed by the number of denoising steps rather than audio length, converging in roughly 8 parallel passes.
Anthropic resumed global access to Claude Fable 5 on July 1, 2026, following the lift of US export controls on June 30 that had forced a suspension since June 12. The pause was triggered by an Amazon report showing a prompt that could bypass safeguards to identify software vulnerabilities and produce exploit code, though Anthropic demonstrated the technique was reproducible on many other models including GPT-5.5 and earlier Claude versions. A newly trained safety classifier now blocks that specific technique in over 99% of cases, rerouting blocked requests to Claude Opus 4.8 instead of refusing them outright. Anthropic also proposed a four-criterion jailbreak severity framework and plans immediate mitigations for the most severe class. During the pause, Zhipu AI released GLM-5.2, a 750B-parameter open-weight model with 40B active per token and a cost of $1.40/$4.40 per million tokens, offering a much cheaper alternative. Fable 5 is priced at $10 per million input tokens and $50 per million output tokens, with access included for Pro, Max, Team, and select Enterprise plans up to 50% of weekly limits through July 7.
NVIDIA released Nemotron-Labs-TwoTower under open weights, a discrete diffusion language model that uses a frozen autoregressive context tower (Nemotron-3-Nano-30B-A3B) and a separately trained denoiser tower. The model retains 98.7% of the AR baseline's aggregate benchmark quality while delivering 2.42× higher wall-clock generation throughput (γ=0.8, block size 16 on 2×H100). The denoiser was trained on ~2.1T tokens, only a fraction of the backbone’s 25T-token pretraining. A single checkpoint provides three generation modes: full mask diffusion, mock-AR, and standard AR decoding. The two-tower architecture uses layer-aligned cross-attention and Mamba-2 state seeding to preserve context representation across diffusion steps.
Anthropic released Claude Sonnet 5 on June 30, 2026, calling it its most agentic Sonnet model. It outperforms Sonnet 4.6 on every published benchmark, including SWE-bench Pro (63.2% vs 58.1%), OSWorld-Verified (81.2% vs 78.5%), and Humanity’s Last Exam with tools (57.4% vs 46.8%), and nearly matches Opus 4.8 on several evals while edging it on GDPval-AA v2 (1618 vs 1615). Introductory pricing is $2/$10 per million input/output tokens until August 31, 2026, then $3/$15, undercutting Opus 4.8’s $5/$25. The model supports effort levels and shows its best cost-performance value at low and medium effort, but at xhigh effort it can cost more than Opus for similar quality. Sonnet 5 uses an updated tokenizer that may increase token counts by up to 1.35×, and its cyber capabilities are intentionally kept low, with Opus remaining the recommended model for accuracy-critical tasks.
Meta AI has released Brain2Qwerty v2, a non-invasive brain-to-text decoder that reconstructs typed sentences from magnetoencephalography (MEG) signals without surgery. The system achieves 61% average word accuracy (39% WER), a major jump from the 8% baseline of prior non-invasive methods; the best participant reached 78% accuracy with over half of sentences having at most one word error. The pipeline combines a convolutional encoder, transformer, and character-level language model, and it leverages fine-tuned large language models to incorporate semantic context. Accuracy improves log-linearly with the amount of MEG recording data. Meta has publicly released the full training code for both v1 and v2 under the CC BY-NC 4.0 license. This is a research release, tested on 9 healthy volunteers in a controlled lab setting; practical use is limited by the need for a magnetically shielded room and a still participant.
This tutorial presents a reproducible Colab workflow for the Fable 5 Traces coding-agent dataset (Glint-Research/Fable-5-traces) from Hugging Face. It manually downloads the merged JSONL trace file, then builds pure-Python utilities to extract tool names, arguments, and text payloads from agent outputs. An audit of the dataset includes detection and redaction of potential secret-like patterns (API keys, tokens) and distribution plots for output types, tools, source roots, and text lengths. Safe no‑CoT chat/SFT exports are created with train/validation/test splits by converting each trace into system-user-assistant messages. A pure NumPy TF-IDF + SVD projection visualises context embeddings, and two pure-Python Naive Bayes classifiers are trained to predict assistant output type and tool name from context, with metrics and top tokens saved. The workflow outputs analysis indices, classifier reports, and a keyword search demo, all without requiring fragile scientific libraries.