Thinking Machines Lab released Inkling, a 975B-total / 41B-active MoE that natively handles text, image, and audio inputs. Architecture: 66-layer decoder-only transformer, 6/256+2 shared expert routing, hybrid local/global attention, hierarchical patch encoder for images, discrete token encoding for audio. Benchmarks place it competitively on AIME 2026 (97.1%), GPQA Diamond (87.2%), SWEBench Verified (77.6%), MCP Atlas (74.1%), and IFBench (79.8%). Eval comparisons run against Kimi K2.5/K2.6, GLM 5.2, DeepSeek V4 Pro, Gemini 3.1 Pro, Claude Fable 5, and GPT-5.6 Sol. Available via SGLang, vLLM, and HuggingChat; 882 HF likes and 7,870 downloads in its first 24 hours suggests meaningful adoption interest.
Hume AI published Real World VoiceEQ, a benchmark covering 40+ proprietary and open-source voice models across 15+ evaluation dimensions and 60+ metrics spanning ASR, TTS, Speech-to-Speech, and Speech Understanding — grounded in 1M+ human ratings (785k TTS, 48k STS). Key finding: no single model ranked in the top five across all eight TTS capability groups; speech models have become better at speaking than listening, often ignoring paralinguistic cues like tone, hesitation, and emphasis that humans rely on. The benchmark also found that LLM-based auto-evaluation degrades significantly for subjective voice quality judgments, suggesting SLMs can't yet replace human raters for open-ended voice assessment.
OpenMOSS-Team's MOSS-Transcribe-Diarize is trending on HuggingFace as an audio-text-to-text model combining speech transcription and speaker diarization in a single Transformer-based model. 83,160 downloads and 242 likes in its first days suggest genuine demand for an open, integrated ASR+diarization pipeline — a category previously dominated by proprietary APIs or awkward two-step open setups (Whisper + pyannote). Worth tracking as a practitioner alternative for meeting transcription and voice agent memory pipelines.
Google launched Gemini Spark — a proactive agentic mode that works on tasks autonomously on the user's behalf (local file management, multi-step projects) — initially to trusted testers then rolling to Google AI Ultra subscribers ($100/month) in the US via the macOS Gemini app. Spark represents Google's first 24/7 cloud agent surface that works server-side without requiring the device to stay online, directly competing with Claude Cowork and ChatGPT Work.
Qwen3 TTS achieves real-time local inference with notably expressive output, integrated into the open-source Persona Engine project (ASR→LLM→TTS pipeline with lip-synced avatar). The author positions it as a meaningful step up from prior local TTS options like Sesame for latency-sensitive, fully offline deployments.
Side-by-side comparison showing GPT Image 2 struggles with photorealistic nature scenes, producing a recognizable artifacting pattern absent in its predecessor. Three images from the same prompt illustrate the regression, flagging a quality tradeoff in the new model for natural/outdoor imagery.
Qwen3.6-35B-A3B is a sparse MoE model with 35B total and only 3B active parameters, released under Apache 2.0. Claims agentic coding performance on par with models 10× its active size, with both multimodal thinking and non-thinking modes. Efficient active-parameter footprint makes it practical for inference on constrained hardware.
🔶 Anthropic Apr 16
⭐ Editor's Pick
Anthropic's official Claude Opus 4.7 GA post confirms same pricing as 4.6, image resolution raised to 2,576px long edge (~3.75 MP, 3× prior), and a new xhigh effort tier. Coding benchmarks: +13% task resolution on internal 93-task harness, 70% on CursorBench (vs. 58%), 98.5% on XBOW visual-acuity (vs. 54.5%). First model shipped with real-time cyber safeguards derived from the restricted Mythos Preview testbed.
Google DeepMind releases Gemini 3.1 Flash TTS, a text-to-speech model featuring granular audio tags for fine-grained control over expressiveness, prosody, and delivery style. The tag-based control interface is the differentiating design choice, enabling more directed audio generation compared to prompt-only TTS systems.
Meta Superintelligence Labs' first model, Muse Spark, is a small, fast proprietary model with native multimodal perception and multi-agent parallel subagent execution—a sharp departure from Meta's Llama open-source strategy. Led by Alexandr Wang, it powers the revamped Meta AI app with Instant and Thinking modes and is rolling out across WhatsApp, Instagram, Facebook, Messenger, and Ray-Ban glasses. API access is restricted to select partners only.