alphasignal depth anything kimi k3 open weights

Why this is in the vault

This issue covers three distinct threads worth tracking: an open-weight coding model (Kimi K3) that benchmarks 2.3x cheaper than Fable on real agent tasks, a local 3D scene reconstruction tool (depth-anything.cpp) with a flat C/Go/Rust API, and a single-pass PDF OCR model (Unlimited-OCR) running fully local on 8GB VRAM. The Kimi K3 vs Fable cost/speed tradeoff is directly actionable for any agentic workflow optimization. The intro thematic — weight ownership and interpretability converging into "who controls intelligence" — is a signal worth watching as open weights become competitive with closed frontier models.

Mapping against Ray Data Co

Kimi K3 vs Fable agent cost tradeoff is live data for RDCO's agentic infrastructure: The Cline team ran both models against a real bug in their own repo. Both fixed it. Fable finished in 3.5 min (18 tool calls, $2.13). Kimi K3 took 12 min (34 tool calls, $0.92). For interactive sessions — code review, vault ops, live coaching — Fable still wins on throughput. For batch/offline agent work (investing thesis drafts, document ingestion, non-interactive vault processing), Kimi K3's 2.3x cost advantage is a meaningful lever worth evaluating. Kimi K3 is open-weight (self-hostable), which also matters for any workloads where data stays on-prem.

Secondary signal: Unlimited-OCR (MIT licensed, 3B params) reads full multi-page PDFs in a single pass with flat memory usage — relevant for any document-heavy data engineering pipeline at phData or RDCO.

⚠️ Sponsorship

Three sponsors in this issue — editorial independence should be weighted accordingly:

• Tiger Data — "One Postgres for AI Data, Now on AWS" — embeddings, model metrics, time-series ingestion; AWS Marketplace CTA. Full sponsor block.
• Agora — "Production-grade infrastructure for conversational AI" — voice AI on SDRTN® network, 80B+ minutes/month, 200+ countries; full sponsor block with "START BUILDING" CTA.
• Bright Data — Inline signal slot (#2 in Signals): petabyte-scale VLA training video by prompt. Sponsor disclosure present.

Sponsor claims are not independently verified.

Issue contents

Intro frame

Lior Alexander opens with a thematic pairing: Mira Murati ships Inkling (975B open-weight model), and researchers discover models reason inside filler tokens hidden from their own output. Frame: "the black box is cracking open from both ends" — weight ownership and internal interpretability are becoming the same question about who controls intelligence and whether it's legible.

Top Repo 1: depth-anything.cpp

• C++ port of ByteDance's Depth Anything 3 by the LocalAI team
• Turns any phone video into a walkable 3D point cloud at 60fps, fully local (no cloud, no subscription)
• Single image/video in → depth map, camera position/angle in 3D space, point cloud exportable as GLB, PLY, or COLMAP, plus sky mask and confidence scores
• 1.3x faster than PyTorch on CPU, half the memory; smallest model is 99MB
• Flat API for Go, Rust, and C embedding — drop-in for non-Python stacks
• 7,272 likes at time of send

Top Model: Unlimited-OCR (Baidu)

• 3B parameter open-source OCR model; reads 100-page PDFs in a single pass
• Key mechanism: Reference Sliding Window Attention (R-SWA) — attends to source + last few written tokens, so memory stays flat regardless of document length (vs. per-page stitching in traditional OCR)
• 93.23% on OmniDocBench — beats prior best by 6+ points
• Runs on single GPU with 8GB+ VRAM; fully local, zero API cost
• Outputs: tables as HTML, equations as LaTeX, full layout preserved
• Compatible with Transformers, vLLM, SGLang, Ollama, Docker; MIT licensed; available on Hugging Face via model.infer_multi()

Top Repo 2: Kimi K3 vs Fable — Cline live benchmark

• Cline tested both models on a real bug from their own production repo
• Fable: 3.5 min, 18 tool calls, $2.13 — 3.4x faster
• Kimi K3: 12 min, 34 tool calls, $0.92 — 2.3x cheaper
• Kimi uses more tokens because it's RL-trained to double-check work before finishing (think longer, output slower, cost less per token)
• First open-weight model that competes directly with top closed models on real coding agent tasks
• Kimi K3 is 2.8T parameters (Moonshot AI); open-weight, self-hostable
• ClinePass trial at $1.99: npm i -g cline

Signals

1. OpenPlanter (4,172 likes) — maps corporate and political power like Palantir; open-source, free
2. Bright Data ⚠️ sponsor — Train VLAs on real-world video, not synthetic data; petabyte-scale clips by prompt
3. AgentScope (23,456 stars) — open-source personal AI assistant, local or cloud deployment
4. Filler token reasoning (4,088 likes) — adding dots to LLM prompts triggers hidden chain-of-thought reasoning readable from the token stream
5. DeepLoop (1,219 likes) — looped transformer architecture, stable and scalable without adding parameters
6. Steganographic LLM text (3,334 likes) — open-source tool hides secret messages inside AI-generated text

Related

• [[2026-07-13-alphasignal-subagents-math-proof-cycle-cover]] — parallel subagents cracking a 50-year math problem; connects to the compute/cost frontier as open-weight models close the gap
• [[2026-07-07-alphasignal-claude-j-space-interpretability]] — Claude J-space exposes hidden reasoning; rhymes directly with this issue's filler-token thread (models reasoning in hidden token space)