This was the week Meta came back to open source, xAI took the image editing crown, and Alibaba invented a new input modality for video. Here’s what happened August 5–11, 2026.
The headline: Meta’s open-source pivot
Meta made two moves this week that reshuffled the open-weight landscape:
Muse Spark 1.2 (August 5) — Meta’s coding-focused model, co-trained with Muse Code (their first terminal agent). TB 2.1 82.9%, AA Index 54, $1.25/$4.25 standard or a jaw-dropping $0.10/$0.20 on the contributor tier. Trails Opus 5 on every coding benchmark — competes on price, not peak capability.
Muse Glimmer (August 10) — A 30B dense model under Apache 2.0, Meta’s first fully permissive open-source release. Runs on 24GB VRAM with DFlash speculative decoding (3.1× on RTX 5090). Day-0 llama.cpp/vLLM/transformers support. Zuckerberg accompanied it with a 14-page letter on open AI and committed to releasing Spark 1.2 weights “in the coming weeks.”
The signal: distribution matters more than control. Five days from closed API launch to Apache 2.0 sibling.
Image generation: xAI’s big jump
Grok Imagine Image 2.0 (August 7) landed at #2 globally on both the Arena text-to-image (Elo 1,320) and image edit (Elo 1,439) leaderboards — behind only GPT-Image-2. The previous Quality Mode sat mid-table on both.
The editing toolkit is real: magic wand, segmentation, background removal, multi-reference compositing (up to 5 images), smart resize across 9 aspect ratios, and professional templates. API access is “coming soon.”
Video generation: document-to-video arrives
Wan 3.0 (August 6, public beta) extends clips to 30 seconds (2× Wan 2.7) and introduces document and web page inputs — PDFs, PowerPoints, spreadsheets, and URLs converted directly into video. No other major video model does this. Priced at $0.05–$0.20/sec. The trade-off: unlike Wan 2.7’s open weights, Wan 3.0 is closed and cloud-only.
MAGI-2 Preview (August 5) is Sand.ai’s 114B MoE (6B active) video model under Apache 2.0 — the first open-source 100B-scale video generation model. Generates 10-second clips with synchronized audio. Requires 8 Hopper GPUs. Research preview, not production, but the complete weights + code + report are available.
OpenAI updates: Sol slider, Luna for free, Cyber for defenders
GPT-5.6 Sol got reliability improvements and a new reasoning depth slider on August 6. GPT-5.6 Luna became the default model for free users with unlimited text chats and a new Think button for deeper reasoning.
GPT-5.6-Cyber (August 10) is a Sol fine-tune for cybersecurity: 95% completion on advanced cybersecurity tasks (vs 57.3% from GPT-5.5-Cyber). Limited to vetted security vendors and researchers. Part of OpenAI’s Daybreak program.
Also this week
| What | Who | When | Why it matters |
|---|---|---|---|
| Nemotron 3.5 Lightning | NVIDIA | Aug 11 | Open-weight model optimized for long-running agentic workloads. Ships with NeMo Switchyard for multi-model routing. |
| Qwen 3.8-Max open weights | Alibaba | Aug 12 (tomorrow) | ModelScope countdown confirms the 2.4T model’s weights drop tomorrow. Qwen3.8-27B to follow. |
| NVIDIA Magpie TTS update | NVIDIA | Aug 10 | 364M multilingual TTS adds Arabic, Korean, Brazilian Portuguese. Open weights. |
| webAI TwiL-LM | webAI | Aug 10 | 1.7B/3B formal-logic models that run on an iPhone. Non-commercial license. |
Three trends from this week
-
Meta’s open-source return. Apache 2.0 Glimmer plus a Spark 1.2 open-weight commitment in the same week. After Llama’s custom licenses and Muse’s closed API, this is a genuine strategic shift — driven by the conclusion that distribution matters more than gatekeeping.
-
Local agentic models are a category now. Muse Glimmer (30B, 24GB), Nemotron 3.5 Lightning, TwiL-LM (runs on iPhone) — three independent labs all shipping models designed to run agents on hardware you own. Cloud inference is no longer the only game.
-
Video inputs are expanding beyond text and images. Wan 3.0’s document-to-video is genuinely new. MAGI-2’s 114B MoE at Apache 2.0 makes 100B-scale video generation open for the first time. The video model space is diversifying faster than LLMs did at the same stage.
For model details, see each model’s page linked above. For the broader LLM landscape, see the 2026 LLM guide.