Week 2026-W37

Projects

1. 01

   HF model AI / ML trending #26 ♥ 4,424 ↓ 1,294,550 ✓ read

   deepseek-ai/DeepSeek-V4.1-Flash

   image-text-to-text · transformers

   This is the week's most consequential release: a 552B multimodal MoE with million-token context whose real story is the systems architecture — causal encoder-decoder, FP4 KV caching, and sparse-index reuse compressing persistent KV state to ~890 bytes/token. That combination could reset expectations for long-context, input-heavy multimodal inference.

   Look for Verify the KV-compression and long-context quality claims against independent evals, and check whether the custom serving stack is practical outside DeepSeek's own infrastructure.

   3 min read ·Hugging Face ↗ ·mit

2. 02

   GitHub Neuroscience ★ 16 ✓ read

   suanmiao/fly-self-driving

   A fruit fly connectome (165k neurons, 25.6M synapses) trained to drive a simulated street from pixels. Code, patches, renderer, results.

   A recurrent network hard-constrained to the measured 165k-neuron Drosophila connectome, with only synaptic gains learned, drives a simulated car from pixels with 58/60 collision-free held-out runs — a concrete test of whether real biological wiring supports closed-loop visual control. Shuffled-wiring baselines and released code make it far more than a stunt.

   Look for Single-seed simulation results and how much heavy lifting the learned gains and DAgger do versus the connectome topology itself — check the shuffled-control gap carefully.

   3 min read ·GitHub ↗ ·Python·MIT

3. 03

   HF model Speech / Video trending #11 ♥ 957 ↓ 18,759 ✓ read

   m-a-p/YuE2-3B

   text-to-audio

   An open 3B text-to-song model generating 48 kHz stereo vocals+accompaniment, with a genuinely new interface: inspectable, editable ABC symbolic plans and an agentic revision loop before acoustic synthesis. Local inference on a 24GB GPU makes it the most usable open music generator to date, squarely in the reader's audio interests.

   Look for Compare vocal/mix quality against Suno-class proprietary systems (see WildSongBench P53) and check the license terms for generated audio.

   3 min read ·Hugging Face ↗ ·cc-by-nc-4.0

4. 04

   GitHub Robotics ★ 354 ✓ read

   showlab/Show-Harness

   Just a VLM Agent Can Play Robots

   A clean framing for embodied agents: the VLM emits incremental semantic actions and deterministic embodiment-specific interpreters execute them, shared across Franka, Piper, ManiSkill, and Isaac Lab. The release is unusually complete — zero-shot frontier-VLM control, small-VLM fine-tuning, adapters, data, and browser-based demonstration collection without hardware.

   Look for How far the semantic-action vocabulary stretches beyond pick-and-place style tasks, and real-robot success rates versus simulation-heavy demos.

   3 min read ·GitHub ↗ ·Python·Apache-2.0

5. 05

   HF model AI / ML trending #2 ♥ 2,214 ↓ 8,109 ✓ read

   Edge0/Edge0-35B-A3B-preview

   text-generation · mlx

   Streaming 4-bit MoE experts from SSD with a learned one-step-ahead prerouter and Recover-LoRA adapters puts a 35B sparse model in under 3 GiB active memory at ~15 tok/s on Apple Silicon — a real jump for edge inference, not another quantized checkpoint. The 79.2 vs 83.2 benchmark gap against the FP16 base is a plausible trade.

   Look for Preview-stage, self-reported numbers tied to the MLX framework; check SSD wear/latency behavior and whether the approach ports beyond Apple Silicon (the 8B sibling P26 runs in ~1 GiB).

   3 min read ·Hugging Face ↗ ·apache-2.0

6. 06

   GitHub Speech / Video ★ 83 ✓ read

   Westlake-AGI-Lab/WorldinWorld

   Official implementation of WorldinWorld

   Re-filming ordinary videos from arbitrary camera paths using a frozen 14B video world model plus depth warping — no training or fine-tuning — is a genuinely new capability with runnable code, not just a paper. Bullet-time and free-camera re-cinematography from real footage is a meaningful demonstration of world models as controllable renderers.

   Look for Requires ~86 GB of weights and an 80 GB GPU; check geometric consistency on large camera excursions where depth-warping priors break down.

   3 min read ·GitHub ↗ ·Python

Also notable

• TaoLiveAIGC/TaoMate-H3 HF model ♥ 63

  Streaming joint audio-video generation in 5-second chunks with a reported 10.6× cut in first-playable latency — a real step toward interactive long-form AV generation, though it needs 4–8 Hopper GPUs.

• zgcagi/ZGCM-1 GitHub ★ 120

  A genuinely open end-to-end 7B training recipe (data, FP8/Muon, 256K context) with near-frontier math and tool-search results at that scale — worth inspecting the recipe even if claims need verification.

• eduardzamfir/RTI GitHub ★ 14

  Content-shaped region tokens for pixel-space diffusion transformers give up to 2.59× speedup with one checkpoint across token budgets — a clean, reproducible efficiency technique with adapters released.

• kevinzakka/mjbatch GitHub ★ 488

  kevinzakka's mjbatch runs thousands of MuJoCo sims in parallel on CPU with per-sim model parameters — likely to be widely adopted for RL, MPC, and system identification.

• nftechie/flm GitHub ★ 68

  Couples a frozen LM to the full MaleCNS fly connectome as a fixed compute substrate; notably honest that the matched control slightly beat the fly topology — a useful negative result alongside P22.

• tencent/AuK HF model ♥ 345

  Tencent's 1.5B unified instruction-driven speech model spanning TTS, editing, paralinguistic control, enhancement, and separation is an interesting direction, but the model card lacks performance evidence (four-step AuK-Flash variant is P34).

• XiaomiRobotics/Xiaomi-Robotics-U0-4B HF model ♥ 11

  Xiaomi's open autoregressive model unifying image generation with robot-centric view transfer and embodied observation synthesis in one token space is relevant to robot world models, though evidence is thin.

• m-a-p/SheetSage2 HF model ♥ 72

  SheetSage2 transcribes full recordings into editable ABC lead sheets, MIDI, chords, and structure with a practical export pipeline — a useful companion to the YuE2 ecosystem.