Open weights pile on as Anthropic digs in

A busy day for the open camp: DeepSeek shipped an MIT-licensed inference speedup, DeepReinforce dropped a self-scaffolding coding model, and Meituan claims a 1.6T-parameter model trained start-to-finish on Chinese silicon. Meanwhile Dario Amodei spent his time warning Congress that open source is dangerous, even as Washington kept Anthropic's own cyber model on a short leash. The center of gravity in shipped, usable AI keeps drifting toward open weights and the serving layer.

DeepSeek's DSpark claims 60-85% faster decoding, MIT-licensed

DeepSeek open-sourced DSpark, a speculative-decoding framework, plus DeepSpec, a codebase for training and evaluating draft models, under the MIT license. It pairs semi-autoregressive drafting (a parallel backbone with a lightweight sequential head) with confidence-scheduled verification that trims low-confidence draft tokens under heavy serving load. Reported per-user generation speedups are 60-85% for V4-Flash and 57-78% for V4-Pro over the prior MTP-1 baseline; offline tests show accepted-length gains carry over to Qwen3 and Gemma4 targets. Early community benchmarks of single-stream V4-Flash land near the paper's ~2.3x-over-no-spec figure.

Why it matters: Speculative decoding is established, but DSpark ships production-tested numbers, open checkpoints, and a training pipeline you can point at your own open-weight model — assuming you control the serving stack and can stomach the ~38TB target-cache requirement.

Ornith-1.0: open-weight coding models that learn their own scaffold

DeepReinforce released Ornith-1.0, an MIT-licensed family (9B dense plus 35B and 397B MoE) post-trained on top of Gemma 4 and Qwen 3.5, both Apache 2.0. The pitch is self-scaffolding: RL optimizes not just solution rollouts but the agent scaffold that drives them, claiming state-of-the-art open-source results on Terminal-Bench 2.1, SWE-bench, NL2Repo and ClawEval at comparable sizes. All checkpoints expose an OpenAI-compatible endpoint with tool calling and a 256K context; the 9B fits on a single 80GB GPU and there are GGUF builds for llama.cpp and Ollama.

Why it matters: Another credible open agentic-coding stack that runs locally and plugs into existing harnesses (OpenHands, OpenCode) — Simon Willison reports it ran a multi-tool agent loop competently over a real codebase.

Meituan's LongCat-2.0: 1.6T params trained entirely on domestic chips

Meituan open-sourced LongCat-2.0, a 1.6-trillion-parameter model with a 1M-token context window, and claims it is the first trillion-parameter model to complete both pre-training and inference on a ~50,000-card domestic cluster of AI ASIC superpods. That goes a step beyond DeepSeek-V4-Pro, which Meituan says used home-grown chips only for inference. Pre-training is the far more compute-intensive phase, making the claim notable if it holds up.

Why it matters: If verified, it signals Chinese accelerators can handle frontier-scale training, not just inference — eroding one of the assumptions behind US export controls.

Amodei warns Congress on open source as Washington leashes Anthropic's cyber model

Dario Amodei used a June 28 congressional hearing to argue open-source models could take us somewhere dangerous, claiming you cannot see inside open models and that they ultimately must be cloud-hosted — assertions the local-model community loudly disputes, given open weights, fine-tunes and at-home inference are the entire point. In parallel, the administration allowed only a limited release of Anthropic's cyber-capable model, part of broader US moves to restrict frontier releases from Anthropic and OpenAI.

Why it matters: The framing fight matters for policy: definitions of what's safe to release shape future export and licensing rules, and Anthropic is simultaneously the loudest anti-open voice and a target of the same restrictions.

Base44 trains its own model to escape the frontier-API bill

Wix-owned vibe-coding platform Base44 began rolling out Base1, an in-house LLM trained on a dataset built from tens of millions of real user interactions. Founder Maor Shlomo frames it as a play for defensibility and margin — owning the stack to optimize latency, cost and efficiency, and eventually beat general frontier models like Opus on app-building tasks. Skeptics note Harvey abandoned its own-model plans, and frontier labs (Claude Code, Cursor) are encroaching on the same turf.

Why it matters: It's a concrete data point in the build-vs-buy debate: as inference costs bite, applied AI companies with enough usage data are weighing vertical integration over renting someone else's frontier model.

Cursor ships a phone app for driving coding agents

Cursor launched Cursor Mobile, letting users spin up new coding agents or steer desktop-initiated ones from their phone, tying into the agent-centric Cursor 2.0 model. It follows similar mobile apps from Anthropic and OpenAI, part of a broader shift from editing code toward supervising code-writing agents — Anthropic's Boris Cherny says most of his coding is now on his phone.

Why it matters: Mobile-first agent oversight signals where the coding workflow is heading: less time in the editor, more time reviewing and approving autonomous agents from anywhere.

Chip geopolitics: Korea's $1T bet, Taiwan raids Super Micro

South Korea committed $1 trillion across memory-chip production, AI data centers and humanoid robots, with President Lee calling semiconductors, physical AI and data centers the triple axis for a great leap forward. The same day, Taiwanese prosecutors raided Super Micro offices and partner firms over alleged smuggling of Nvidia AI chips into China; Super Micro's stock fell 8% and a co-founder was reportedly indicted.

Why it matters: The hardware supply chain is now an explicit instrument of state policy — both massive subsidies and criminal enforcement — and that volatility flows straight through to GPU and memory prices developers pay.

Also worth a look

Browse previous days →