The interesting action isn’t just Nemotron‑scale models and AGI timelines; it’s that Gemma‑class multimodal models now run on a 16GB laptop while Google signs a $920M‑per‑month GPU lease. Cheap Chinese and open‑weight models like DeepSeek, Qwen, GLM, and MiniMax M3 are quietly seizing the cost–performance frontier, as agents like Scout and Claude‑powered toolchains show that state, memory, and reliability—not pure IQ—are where things are actually breaking.
The gap between frontier marketing and what’s practically deployable is widening, and the most important optimization knobs now live in routing, compression, and governance rather than just picking “the best model.”
Key Events
/SpaceX will provide Google about 110,000 NVIDIA GPUs under a $920M‑per‑month compute deal running through 2029.
/Gemma 4 12B launched as an Apache‑licensed multimodal model with a 256K context window that runs on 16GB laptops, plus QAT checkpoints that cut memory use by ~3x.
/DeepSeek V4 Flash is reported to be ~10x cheaper than Opus 4.8 at similar performance, as DeepSeek targets a $7B funding round.
/NVIDIA released Nemotron 3 Ultra, a 550B‑parameter hybrid MoE model with 55B active parameters and a 1M‑token context, marketed as 5x faster and 30% cheaper than rivals.
/Anthropic says Claude now writes over 80% of its own production code and that Opus 4.8 set a new state of the art on ARC‑AGI‑3 at ~1.5%.
Report
Everyone is staring at ARC‑AGI and 550‑billion‑parameter monsters; the more interesting move is that near‑frontier models now run on a 16GB laptop.
At the same time, Google is committing about $920M per month just to rent GPUs from SpaceX.
the split‑screen compute economy
SpaceX is supplying roughly 110,000 NVIDIA GPUs to Google under a deal worth $920M per month, with the arrangement totaling around $30B through 2029.
Google and Anthropic together expect to spend about $26B per year on these GPU resources. Alphabet is separately raising $80B for additional AI infrastructure, signaling a massive capital push into centralized compute.
In parallel, Gemma 4 12B runs on a 16GB laptop, and its QAT variant hits ~120 tok/s on a 12GB GPU while using roughly 3x less memory with near‑original performance.
Huawei’s KVarN compresses KV cache by 3–5x while matching higher‑precision baselines, and can be enabled in vLLM with a single flag.
china‑linked open models are quietly seizing the cost frontier
DeepSeek V4 Flash is reported to be about 10x cheaper than Opus 4.8 while offering similar performance, and DeepSeek is raising $7B to scale further.
DeepSeek scored 103/120 on the Putnam‑style exam, putting it in elite math territory compared with many human competitors. Qwen 3.6‑27B outperforms Gemma 4 12B in five of eight benchmarks despite a smaller footprint, and Qwen 3.6‑35B is praised for speed and KV‑cache efficiency.
GLM 5.1 powers a hybrid legal agent that beats Claude Opus 4.7 on all‑pass rate (18% vs 14%), and is described as the leading open‑source model in its class.
Open‑source AI has grown 500% in two months, while some benchmarks show Kimi K2.6 and GLM 5.1 outperforming NVIDIA’s Nemotron 3 Ultra on coding despite its 550B parameters.
agents are bottlenecked on state, not raw intelligence
OpenClaw has accumulated over 5,300 community skills and inspired Microsoft’s Scout agent, which runs continuously with ~50 tools and can assist users even when they are offline.
Early users report Scout falsely marking tasks as complete and crashing mid‑execution, prompting quick bug fixes and internal concern about making it “addictive.” LangGraph is becoming a default for customer‑support agents, but developers complain about opaque state management, hard‑to‑debug multi‑agent flows, and a “black box” feel.
Hermes Desktop offers persistent workflows and a memory system that survives rebuilds, yet users struggle with installation issues and the complexity of configuring agents.
MCP tooling shows the same pattern: GitHub’s MCP server can burn 17k tokens before a real user query, while CostAffective MCP and ActionFence exist mainly to rein in exploration loops and spend.
self‑improving code loops vs actual cognition
Anthropic reports that more than 80% of code merged into its production codebase is now authored by Claude, and frames this as a potential recursive self‑improvement dynamic.
The company’s Mythos line achieved a 52x speedup in training tasks and task horizons of three to four hours, with Anthropic calling for a global freeze on AI development as Claude writes most new code.
Opus 4.8 set a new state of the art on ARC‑AGI‑3 with a score around 1.5% at roughly $10,000 per run, and some commentators describe this as proto‑AGI territory.
OpenAI and Anthropic both say they see early signs of recursive self‑improvement, while Sakana AI has set up a dedicated RSI lab and some forecasters place plausible RSI in the 2027–2028 window.
Yet top frontier models like GPT‑4o, Claude 3.5, and Gemini 2.5 still completely fail the Stroop attention test, and users continue to report that Claude and peers struggle with fundamental reasoning tasks and require extensive guidance.
open multimodal + voice are commoditizing, video is still the hard boss
Gemma 4 12B is a unified, encoder‑free multimodal model that accepts text, images, and audio directly with a 256K context window, runs on 16GB RAM laptops, and ships under Apache 2.0.
MiniMax M3 combines coding, multimodal inputs, and a 1M‑token context window, scoring 59% on SWE‑Bench Pro and 66% on Terminal Bench 2.1 as the first open‑weights model with this combo.
Ideogram 4.0, a 9.3B open‑weight model, sits at the top of Image Arena, can generate true‑2K and up to 10‑megapixel images, and is tuned for layout‑driven editing via ComfyUI and Z‑Image.
Cosmos 3 is an open omnimodel that handles language, images, video, audio, and actions, ranking #1 for both text‑to‑image and image‑to‑video tasks.
Meanwhile, open TTS models like dots.tts (2B, zero‑shot cloning), Alibaba’s Fun‑Realtime‑TTS (top of Speech Arena), and Miso One (8B, ~110ms latency) make high‑quality voice interfaces available without frontier APIs, while video remains explicitly called out as a challenging modality and WAN 2.2 users report slower generations and audio sync issues.
coding assistants are getting more capable and more expensive at the same time
Anthropic says Claude Code now writes over 80% of its own production code, and that engineers are shipping eight times as much code per quarter compared with earlier years.
Claude Mythos extends this with 3–4‑hour task horizons and reports of 52x faster training code generation than human efforts. At the same time, Uber blew through its AI budget in Q1 and capped some AI coding tools at $1,500 per month, while Microsoft reportedly pulled Claude licenses from developers due to cost.
GitHub Copilot’s shift to metered billing has led to cases where a $29 subscription ballooned into bills in the hundreds of dollars, and a VS Code zero‑day allowed attackers to steal GitHub tokens and launch a supply‑chain attack on over 100 repositories.
Developers increasingly report that AI‑generated code is driving a noticeable rise in bugs and technical debt, even as some teams claim 7.4x–17.3x boosts in code output from local and remote coding agents.
What This Means
The center of gravity is drifting away from a single dominant frontier model toward a messy ecosystem where cheap Chinese open‑weights, laptop‑class multimodal models, and brittle but powerful agents all coexist—and where the real constraint is no longer raw model size but how much state, memory, cost, and risk you can actually manage.
On Watch
/Huawei’s KVarN KV‑cache quantization already delivers 3–5× compression with q8_0‑level accuracy, but incompatibility with unified KV cache systems is generating active community workarounds that could reshape which inference engines win.
/Multi‑token prediction (MTP) is showing 3–5× decode speedups on Gemma 4 and other models but also triggering VRAM spikes and prompt‑processing slowdowns on some Vulkan setups, making its real‑world trade‑offs a live question.
/Apple’s upcoming Siri 2.0, running on NVIDIA Blackwell GPUs with Gemini integration and opening iMessage to AI agents, could abruptly make Apple a serious player in the assistant/agent ecosystem if the long‑derided Siri UX actually improves.
Interesting
/MAI-Thinking-1, a new text foundation model, has 35B active parameters and achieved 97% on AIME 2025, excelling in reasoning tasks.
/An advanced AI model named o1 has outperformed physicians in evaluating emergency department conditions, showcasing AI's potential in healthcare.
/Anthropic has confidentially submitted a draft S-1 statement for a potential IPO, with an annualized revenue run rate nearing $50B.
/Cerebras' achievement of creating a single chip from an entire 46,225mm² wafer demonstrates near 100% usable wafers.
/InstinctRazor-Qwen3.5-122B-A10B-GGUF is a 122B MoE model that operates with only 8 GB of active GPU VRAM by utilizing CPU for some experts.
We processed 10,000+ comments and posts to generate this report.
AI-generated content. Verify critical information independently.
/SpaceX will provide Google about 110,000 NVIDIA GPUs under a $920M‑per‑month compute deal running through 2029.
/Gemma 4 12B launched as an Apache‑licensed multimodal model with a 256K context window that runs on 16GB laptops, plus QAT checkpoints that cut memory use by ~3x.
/DeepSeek V4 Flash is reported to be ~10x cheaper than Opus 4.8 at similar performance, as DeepSeek targets a $7B funding round.
/NVIDIA released Nemotron 3 Ultra, a 550B‑parameter hybrid MoE model with 55B active parameters and a 1M‑token context, marketed as 5x faster and 30% cheaper than rivals.
/Anthropic says Claude now writes over 80% of its own production code and that Opus 4.8 set a new state of the art on ARC‑AGI‑3 at ~1.5%.
On Watch
/Huawei’s KVarN KV‑cache quantization already delivers 3–5× compression with q8_0‑level accuracy, but incompatibility with unified KV cache systems is generating active community workarounds that could reshape which inference engines win.
/Multi‑token prediction (MTP) is showing 3–5× decode speedups on Gemma 4 and other models but also triggering VRAM spikes and prompt‑processing slowdowns on some Vulkan setups, making its real‑world trade‑offs a live question.
/Apple’s upcoming Siri 2.0, running on NVIDIA Blackwell GPUs with Gemini integration and opening iMessage to AI agents, could abruptly make Apple a serious player in the assistant/agent ecosystem if the long‑derided Siri UX actually improves.
Interesting
/MAI-Thinking-1, a new text foundation model, has 35B active parameters and achieved 97% on AIME 2025, excelling in reasoning tasks.
/An advanced AI model named o1 has outperformed physicians in evaluating emergency department conditions, showcasing AI's potential in healthcare.
/Anthropic has confidentially submitted a draft S-1 statement for a potential IPO, with an annualized revenue run rate nearing $50B.
/Cerebras' achievement of creating a single chip from an entire 46,225mm² wafer demonstrates near 100% usable wafers.
/InstinctRazor-Qwen3.5-122B-A10B-GGUF is a 122B MoE model that operates with only 8 GB of active GPU VRAM by utilizing CPU for some experts.