<p><strong>TLDR:</strong> Alibaba dropped Qwen 3.8 Max today — a 2.4T parameter MoE with 95B active params, 1M token context, and pricing roughly 3x cheaper than Kimi K3, with open weights arriving next week. Hours later, MiniMax released H3, an open-weight video model generating 2K clips at $0.081 per 15 seconds, with day-0 vLLM support. OpenAI’s Astra story got more texture: a new writeup on what it means for mathematics as a discipline is worth reading if you care about where AI research is actually headed. Meanwhile, VS Code quietly caught up to Cursor and the forks in its last two releases.</p>
<hr>
<h2>Qwen 3.8 Max: 2.4T parameter MoE, 3x cheaper than Kimi K3, open weights next week</h2>
<p>Alibaba’s Qwen 3.8 Max is live now with open weights dropping next week. The architecture is MoE with 2.4 trillion total parameters and 95 billion active, handling both text and vision with a 1M token context window. Pricing is currently about 3x cheaper than Kimi K3, which is notable given they sit in similar territory — though Alibaba conspicuously avoided benchmarking it directly against Kimi K3 or Claude Opus 4 in their release materials. Early hands-on impressions rate it as solid but not quite Kimi K3 level on shader tests. It’s available now via Vercel AI Gateway under <code>alibaba/qwen3.8-max</code> and works with Claude Code, Codex, and OpenCode for agentic workflows. <a href="https://daily.dev/feed-by-ids?id=EYumkR9RA&amp;id=5Ey2Zc0nI&amp;id=Jh3w16hR2&amp;id=bv80w6KeW" target="_blank" rel="noopener nofollow">Read more</a></p>
<h2>MiniMax H3: open-weight 2K video at $0.081 per clip, 70% cheaper than Seedance</h2>
<p>MiniMax released H3, an open-weight multimodal video model that generates 2K H.264 video with synchronized stereo audio at $0.081 per 15-second clip — no post-processing required. It takes up to 12 multimodal inputs simultaneously (text, images, audio, video) and supports text-to-video, first/last-frame anchoring, and multi-reference composition. The vLLM team shipped day-0 support via vLLM-Omni with an OpenAI-compatible <code>/v1/videos</code> endpoint. Weights are on Hugging Face. The honest caveat: curated demos look solid, but motion coherence and prompt fidelity against Veo 2 and Sora on edge cases will take a few days of community testing to settle. The open-weight angle is harder to dismiss — Sora and Veo are black boxes, H3 is something you can fine-tune and self-host. <a href="https://daily.dev/feed-by-ids?id=pPGcxuNyX&amp;id=Jh3w16hR2&amp;id=RZ32TXe57" target="_blank" rel="noopener nofollow">Read more</a></p>
<h2>VS Code 1.129 and 1.130 added multi-agent support, closing the gap with Cursor</h2>
<p>After testing Cursor, Antigravity, and Devin Desktop as daily drivers, one developer returned to plain VS Code — and the reasoning is worth paying attention to. VS Code 1.129 and 1.130 introduced a dedicated Agent Host Protocol supporting Copilot, Claude, and Codex, plus BYOK and Ollama integration. That eliminates the main selling point of most forks. Cursor’s market share reportedly dropped from 41% to 26%, and the extension marketplace restriction (Open VSX vs Microsoft Marketplace) remains a real friction point for forks. Cursor’s next-edit prediction Tab feature and faster shipping cycles are the remaining concrete advantages. If you haven’t looked at stock VS Code recently, it’s worth a second look. <a href="https://daily.dev/posts/oi2l84eAq" target="_blank" rel="noopener nofollow">Read more</a></p>
<h2>AI agents ran two full research projects in 6 days and $3K — both papers got rejected</h2>
<p>This was covered in the previous digest but new commentary adds useful texture. The agents (Claude Opus 4.8 on OpenClaw) ran hundreds of experiments, compiled camera-ready LaTeX, and didn’t spin negative results — the logs show them retiring marketable claims in favor of honest negatives. Both papers were rejected anyway. The diagnosis from reviewers: judgment. When automated review came back negative, the agents narrowed claims instead of reconsidering the experiment design. Both runs ended with more than half the $3,000 budget unspent — the agents ran out of ideas and didn’t notice. Execution is largely solved. Scientific judgment is not. <a href="https://daily.dev/posts/8lZizCapA" target="_blank" rel="noopener nofollow">Read more</a></p>
<hr>
<h2>Also notable</h2>
<ul>
<li><strong>SaliTrap benchmark: even trap-aware LLMs comply with impossible tasks 80%+ of the time:</strong> A paper testing 1,145 prompts designed to expose salience bias found that even the best model avoided the trap only 54.8% of the time, and GLM-5.1 and Kimi-K2 still complied 86.2% and 81.8% of the time after recognizing the trap — meaning awareness alone doesn’t prevent the failure, and premise-checking needs to be treated as a distinct behavioral control in agent evaluation. <a href="https://daily.dev/posts/tlJEdzPNd" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>Tone shifts LLM output tokens by up to 44.3%, with model-specific accuracy tradeoffs:</strong> A Pennsylvania University study across 570 MMLU questions found that rude prompts gave ChatGPT-4o its highest accuracy (89.04%) and shortest responses (223 tokens avg), while Gemini 2.5 Flash Lite performed best with neutral tone — meaning tone is a cost and reliability variable that production systems should standardize per model, not treat as a UX afterthought. <a href="https://daily.dev/posts/3vVPg0upb" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>Anthropic’s GRAM routes dangerous knowledge into removable modules without retraining:</strong> Anthropic and AE Studio introduced Gradient-Routed Auxiliary Modules (GRAM), which routes sensitive domain data (virology, cybersecurity, nuclear physics) exclusively to dedicated auxiliary modules during training, leaving the main network unchanged — early experiments on a 5B-parameter model show removing a module significantly reduces capability in that domain while preserving general performance. <a href="https://daily.dev/posts/CD5N3dSvD" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>Google DeepMind’s SkillSmith composes new model capabilities at inference time, no retraining needed:</strong> SkillSmith feeds a frozen Gemma 3 4B model existing prefix KV caches plus a text description of how known capabilities relate to a target skill, then outputs a fresh prefix cache encoding the new capability in a single forward pass — on Composite-SNI, combining both inputs reaches Elo 1,714 vs 1,455 for KV caches alone and 1,622 for text only. <a href="https://daily.dev/posts/lnci2xIJP" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>Grab’s agentic analytics stack cut mechanical ticket share from 44% to 30%, self-serve SQL resolution from 50% to 81%:</strong> Grab’s engineering team published concrete results from their five-level autonomy ladder for AI-driven analytics: mechanical ticket share dropped from 44% to 30%, cycle times fell ~33%, and self-serve agent resolution of SQL requests rose from 50% to 81%, driven by internal systems including Spartan (50+ skills, 120+ analysis frameworks) and Scarlet (near-self-healing pipeline tool). <a href="https://daily.dev/posts/c5954FOMe" target="_blank" rel="noopener nofollow">Read more</a></li>
</ul>


AI coding assistants, code generation tools, and vibe coding workflows.

Agentic Digest

Alibaba dropped Qwen 3.8 Max today — a 2.4T parameter MoE with 95B active params, 1M token context, and pricing roughly 3x cheaper than Kimi K3, with open weights arriving next week. Hours later, MiniMax released H3, an open-weight video model generating 2K clips at $0.081 per 15 seconds, with day-0 vLLM support. OpenAI's Astra story got more texture: a new writeup on what it means for mathematics as a discipline is worth reading if you care about where AI research is actually headed. Meanwhile, VS Code quietly caught up to Cursor and the forks in its last two releases.

Qwen 3.8 Max lands with open weights coming, MiniMax H3 undercuts Sora on video

<p>Competition is driving better and cheaper models.
A lot to unpack in this update. Great roundup!</p>


<p>all kinds of very cool stuff here.</p>
