Alleged 8TB "Claude Fable 5" weight leak from AWS S3 looks fabricated, teardown finds
AccBalanced · x · 2026-10-31
An actor named "localizer" claims to have pulled a 7GB shard of Anthropic's unreleased "Claude Fable 5" weights from a misconfigured AWS S3 bucket tied to Amazon Bedrock — allegedly 640 shards totaling 8TB — and is now selling it, posting safetensors files and hex dumps as proof. The claim remains unverified.
Community teardown points to fabrication:
- One layer per shard (as the header implies) means 640 layers, 4.7T params with 1T active per token, and a KV cache of 2.6MB/token — 2.6TB for a single 1M-token conversation. Not plausible.
- 8-way tensor parallelism gives a normal 80 layers, but then outb's 32768-wide input implies an absurd 262,144-wide full dimension; splitting by experts yields a 30KB header, not 260KB.
- The architecture looks like a collage: DeepSeek-style query compression without MLA's actual KV-cache point, Kimi K2's 384 experts, MT-NLG's width — each oddity breaking another number.
Verdict: a "platypus model," likely fake — the anatomy is stronger evidence than the seller's sloppy numbers.
Related event: Alleged Claude Fable 5 Weight Leak Likely a Hoax(3 posts)→
More from Models
- Personality benchmark radar charts show DeepSeek's shape diverges from other Chinese models — teortaxesTex · 2026-10-31
- Users complain Opus 5.5 feels heavily nerfed — RileyRalmuto · 2026-10-31
- Bug bounty researcher flags AI security platform sessions as potential training data of unpatched exploits — npinto · 2026-10-31
- MOSS AMA Roadmap: Full-Duplex Live Model, Image Turbo, Consumer-GPU Text2Vid — tomByrer · 2026-10-31
- Grok says rumored Gemini 4 Carbon cyber model is wrapped in guardrails, not leaked carelessly — AccBalanced · 2026-10-31
- NVIDIA ships NVFP4 DeepSeek-V4.1-Flash, 81GB smaller with near-identical quality — AccBalanced · 2026-10-31