Models & Releases 3
2.8T MoE, native visual understanding, 1M context; Fireworks calls it "the first frontier open model in the 3-trillion-parameter class"; the catch surfaced fast on HN — at mxfp4 it still needs ~1.5TB of VRAM, so nobody runs this at home, but the median third-party price will reveal what serving a 3T model actually costs, and competition is expected to crater it (GLM dropped ~45% in six weeks) (9.9k + 2.8k likes)
Moonshot's companion release derives 10 atomic perceptual capabilities from how frontier models fail across 42 benchmarks, then builds 3,000 verified questions each answerable by looking alone — no reasoning or outside knowledge (1.5k likes)
OpenAI's full-duplex voice model (already on consumer plans) rolls out to the org tiers in ChatGPT Voice (1.6k likes)
Tools & Demos 3
three infra releases dropped alongside the model: FlashKDA (CUTLASS Kimi-Delta-Attention kernels, 1.72–2.22× prefill speedup over flash-linear-attention on H20), MoonEP (expert-parallel MoE comms), and AgentENV (distributed agent-environment system with snapshot/resume/fork that powered K3's agentic RL) (2.8k + 2.6k likes)
a minimalist evolutionary demo where tiny agents navigate a maze; a palate cleanser on the theme of how little "intelligence" a task actually needs (23 pts, 7 comments)
python-build-standalone gets a fresh look; the relocatable interpreter builds that quietly underpin uv, Rye, and a lot of AI tooling packaging (118 pts, 24 comments)
AI Agents 4
tackles the diversity-vs-verifiability dilemma in self-evolving training by letting skills co-evolve rather than optimizing tasks in isolation; one of the week's top agent-training papers (30↑)
data quality decides model capability, yet there was no unified benchmark for how well LLMs and agentic workflows actually clean and prepare training data; this builds one (42↑)
argues production agents fail less from bad reasoning than from failing to manage what's in their context (histories, huge prompts, tool defs), and reframes it as an architecture-and-lifecycle problem (17↑)
instead of ever-better screenshot perception, feed the agent the underlying program state directly, since a screenshot is a lossy render — a meaningful gain on long-horizon computer-use tasks (18↑)
Industry 3
a concrete data point under the whole open-weights debate: a major US public company moving frontier-tier workloads to open Chinese models and halving its bill (10 pts)
a push to bring Claude to Cognizant's enterprise clients, the same week Anthropic staked out its open-weights position (2.0k+ views)
swyx notes the podcast quietly overtook Gruber, Cal Newport, and TBPN, and hired ReadWriteWeb/The New Stack veteran Richard MacManus as Head of Editorial (84 + 24 likes)
KOL Corner 3
David Ha: "open ecosystems are the foundation of a healthy AI industry… collective intelligence is the future"; the Japanese lab joins the same letter NVIDIA signed, framing open weights as the basis for technological sovereignty (981 + 101 likes)
riffing on Codex/Claude Code building genuinely novel playable demos on demand (a Piranesi city-builder of arches and waterways, a game "about Imminence"), his point: stop cloning the same six example games, and the whole thing is downstream of the Sparks paper's TikZ unicorn (1.2k + 727 likes)
"Been reading the weights of the most powerful open-weights model yet released. Page one starts strong: mostly negatives, a late run of zeros, and one unexpectedly large positive value" — the day's best joke about what "open weights" literally means (2.6k likes)
HF Trending Papers 4
text-only pre-training limits models' grasp of the physical world; this trains multimodal from the start to avoid that ceiling (18↑)
mainstream RL frameworks thread every change through trainer, backend, and rollout glue; Molt restructures so algorithm changes stay local (24↑)
embodied agents can't just eat the internet; they need data coupling observation with physical state and action, organized as a pyramid of sources (15↑)
existing 3D benchmarks score text answers or single-object ops; this tests whether VLM agents can take real actions on complete 3D scenes (4↑)