Open weights squeeze the frontier labs

The day's thread runs from both ends of the open-weight squeeze: over the weekend developers got DeepSeek's 284B V4-Flash running on used ~$6K servers, while "Big Short" investor Steve Eisman warned that cheap Chinese open models could ignite a price war right before the OpenAI and Anthropic IPOs. On the drama side, OpenAI answered Apple's trade-secret suit with a full document dump.

DeepSeek V4-Flash, a frontier reasoner, now runs on commodity home hardware

Over the weekend LocalLLaMA users got the official 284B-total/13B-active V4-Flash-0731 checkpoint (156GB, QAT-native MXFP4) running on used gear: a quad-Xeon DDR4 server plus two RTX 3090s (~$6K all-in) hits 33 tok/s single-stream and up to 68 aggregate, with a spec-decode + Marlin path giving a ~2.6x jump over ik_llama.cpp. Cold prefill is the weakness (a ~9s fixed floor, TTFT stretching to minutes on long fresh prompts), which pins the box to overnight batch work rather than interactive coding. On quality, testers report Q2 quants degrade below Qwen3.6-27B, Q3 is a reliable Qwen3.6-27B replacement, and full precision approaches GLM 5.2.

Why it matters: A quantization-aware, MXFP4-native frontier-class model you can self-host for pennies of electricity changes the build-vs-buy math for teams that need data sovereignty and can tolerate a batch queue.

Eisman warns cheap Chinese open models could ignite an AI price war before the IPOs

On his show, 'Big Short' investor Steve Eisman said that if he ran OpenAI or Anthropic he'd be 'petrified' of a price war. His specific example: Moonshot's open-weight Kimi K3 at $3/M input tokens versus $5 for GPT-5.6 Sol and $10 for Claude Fable 5, with open weights removing the switching cost premium subscriptions depend on. Both labs have filed confidentially with the SEC targeting ~$1T listings. Bloomberg Intelligence cited 988 approved Chinese LLMs, DeepSeek cutting API prices up to 50%, and Baidu cutting 99% earlier this year.

Why it matters: The moat debate now has an IPO clock on it: the pricing power a trillion-dollar valuation assumes is exactly what an open-weight price war erodes, and public investors will price it directly.

OpenAI answers Apple's trade-secret suit with the chat logs

OpenAI published emails and iMessages to rebut Apple's July complaint, which alleges former Apple engineer Chang Liu improperly accessed confidential files after joining OpenAI. The receipts show Apple's outside counsel emailed the wrong person after confusing two Asian last names and claimed a phone call that OpenAI says never happened, and that Apple employees kept texting Liu for internal files after his January 22 departure. As critics note, the messages don't refute Apple's core claim that OpenAI encouraged new hires to bring proprietary information. The case ties to OpenAI's Jony Ive-led io Products hardware push and 400+ ex-Apple staff.

Why it matters: Good theater, but the document dump sidesteps the central allegation; the real fight is over OpenAI poaching Apple hardware talent for its consumer-device ambitions.

How the giant MoEs actually get served: Cloudflare and Baseten open the playbook

Cloudflare detailed the tricks it layers on SGLang to serve Kimi and GLM: FP8 KV cache (raising Kimi K2.6 in-memory context from ~686K to ~1.37M tokens for ~30% lower cost/token), INT4 weight compression for GLM 5.2 (705GB to 421GB, per-GPU 88GB to 52GB, no accuracy loss), and per-page KV-cache integrity checks under 1% overhead. Baseten's Inference Engineering episode covers disaggregated prefill/decode, traffic-specific speculators, and grafting a Kimi vision encoder onto GLM 5.2 by training only the projector, plus why identical weights loop into repeated tokens on one cluster but not another.

Why it matters: The gap between 'generated a token' and a reliable production API is where 20-200% speedups and margins live; both writeups are unusually concrete about the quantization and routing that get you there.

AI as co-author: two teams crack the same quantum-crypto problem three hours apart

MIT's Seyoon Ragavan and a UCSB/UCLA pair independently solved the same 'unclonable encryption' problem using OpenAI's GPT-5.6 Sol Ultra, posting to arXiv within three hours of each other and now weighing a merged paper. Separately, OpenAI detailed the specific claims behind its internal 'Astra' model: proofs of exponential quantum parallel repetition, stronger closest-vector-problem hardness, and results in sphere packing and Ramsey numbers, all still unpublished and unverified.

Why it matters: When every researcher queries the same model, 'independent discovery' and authorship norms start to wobble; and the AI-generated proofs still need human referees before any of it counts.

LM Studio buries its own app to push the Bionic agent

LM Studio has replaced nearly every download link on its site with its new Bionic agentic harness, demoting the original local-model app to a tiny footer link while the core app has seen only two or three minor updates since Bionic launched. Longtime users read it as a quiet deprecation in favor of an agent (with cloud-model upsells) that not everyone wants, and threads are already asking how to migrate to llama.cpp.

Why it matters: One of the most popular local-LLM front-ends may be deprioritizing the very tool that built its reputation, worth watching if it sits in your local stack.

Also worth a look

Browse previous days →