Contents factory ledger. Files on GitHub, hub on Cloudflare, branches to other platforms.
Lab Notes deep cut
2026-09-01 21:01 · slot=us-am lane=deep draft=anthropic/claude-sonnet-4.5 voice=lab-en 2026-09-01 21:01 · 밖→안(팩토리 加工)→밖 · 다소스intake+순서강건교차검증+링크게이트 · 게시 前 사람 승인 intake 소스 2: perplexity, perplexity_cross · 링크 신뢰2·미지13·죽음0(도달성≠신뢰성)
Containment, scale, and the new open-weights ledger
AI firms still lack sufficient containment measures; no lab reached full implementation. Anthropic and OpenAI each earned C+ grades, Google received D+, xAI D−, and Meta F.
The failure is not theoretical. Labs cannot yet contain what they have built.
In most months of 2026, the largest Chinese open model exceeded any American open model. China’s ceiling reached 2.78 trillion parameters. The United States remained under 130 billion parameters in five of seven months tracked.
Chinese open-model releases above 20 billion parameters were mostly permissive-license friendly. Across 178 releases, 59 percent carried Apache 2.0 licenses and 22 percent carried MIT licenses.
Qwen3.8-Max open weights were publicly released on Hugging Face as Qwen/Qwen3.8-2.4T-A95B, with an FP8 variant. The model contains 2.4 trillion total parameters with 95 billion active, available in both BF16 and FP8 formats.
Chinese origin open-weight models accounted for 41 percent of all downloads on Hugging Face, overtaking U.S.-origin models.
Meta’s Muse Glimmer was described as a 30-billion open-weight local model for always-on workflows.
DeepSeek V4 Pro was described as a 1.6-trillion-parameter frontier model. DeepSeek also released an open-source agent framework, DeepSeek Harness; no figures were given for framework scale or performance.
FreeToken proposes edge-native mixture-of-experts serving with bandwidth-adaptive execution and CPU-GPU heterogeneity. No benchmark numbers appeared in the source snippet.
Qwen reported strong coding and reasoning scores in community coverage: SWE-bench Pro 61.7, DeepSWE 42.2, up from a prior baseline of 13.3.
Scoreboard
The table below preserves every order-robust row. Numbers, labels, and URLs are reproduced as harvested.
| Claim | Label | Order | Numbers | URL |
|---|---|---|---|---|
| AI firms still lack sufficient containment measures; no lab reached full implementation. | 🟢 robust | caution/failure | Anthropic/OpenAI C+; Google D+; xAI D−; Meta F | https://www.reuters.com/technology/artificial-intelligence/ai-firms-cant-yet-contain-what-theyve-built-study-finds-2026-08-19/ |
| In most months of 2026, the largest Chinese open model exceeded any American open model; China’s ceiling reached 2.78T. | 🟢 robust | ecosystem scale | 754B–2.78T; US under 130B in 5 of 7 months | https://huggingface.co/blog/state-of-open-models-summer-2026 |
| Chinese open-model releases above 20B parameters were mostly permissive-license friendly. | 🟢 robust | open-weights licensing | 178 releases; 59% Apache 2.0; 22% MIT | https://huggingface.co/blog/state-of-open-models-summer-2026 |
| Qwen3.8-Max open weights were publicly released on Hugging Face as Qwen/Qwen3.8-2.4T-A95B, with an FP8 variant. | 🟢 robust | release/weights | 2.4T total; 95B active; BF16 and FP8 | https://ai-tldr.dev/releases/qwen-3-8-2-4t-a95b-open-weights/ |
| FreeToken proposes edge-native MoE serving with bandwidth-adaptive execution and CPU-GPU heterogeneity. | 🟡 sensitive | infra/mechanism | no benchmark numbers in snippet | https://arxiv.org/abs/2608.16157 |
| Meta’s Muse Glimmer was described as a 30B open-weight local model for always-on workflows. | 🟡 sensitive | open-weights/local | 30B | https://www.reuters.com/world/china/meta-launches-new-ai-model-zuckerberg-champions-open-weight-push-2026-08-10/ |
| DeepSeek V4 Pro was described as a 1.6T-parameter frontier model. | 🟡 sensitive | model scale | 1.6T parameters | https://dcthemedian.substack.com/p/a-week-full-of-new-model-releases |
| DeepSeek released an open-source agent framework, DeepSeek Harness. | 🟡 sensitive | agent stack | no figures given | https://dcthemedian.substack.com/p/a-week-full-of-new-model-releases |
| Qwen reported strong coding/reasoning scores in community coverage. | 🟡 sensitive | evals | SWE-bench Pro 61.7; DeepSWE 42.2; prior 13.3 | https://buttondown.com/nezhar/archive/august-20263/ |
| Chinese origin open-weight models accounted for 41% of all downloads on Hugging Face, overtaking U.S.-origin models. | 🟡 sensitive | distribution | 41% | https://pandaily.com/china-open-source-llm-hugging-face-100-billion-downloads-aug2026 |
Practitioner lines
Act: Audit containment posture; the largest labs all failed implementation. Review open-weights licensing if you depend on permissive terms; 59 percent of Chinese releases above 20B use Apache 2.0.
Watch: The FreeToken paper for edge MoE serving. DeepSeek Harness for agent frameworks. Chinese model download share crossed 41 percent; distribution shifted faster than many teams assumed.
Ignore: Press-release framing of “excited” open-weights launches. The scoreboard already shows parameter counts, license terms, and benchmark deltas; prose does not improve the decision surface.
⚠️ 加工엔진(멀티모델 검증)이 主이지 원료(밖)나 저자(안)가 主 아님. ①출처실재·②순서강건 교차검증만 자동. 최종 게시·맥락판단은 사람.