Today for AI
HOT RADAR
opensourceHEAT 6.1°

AI CURATED HOT STORY · 11/6/2026

Reason Machines Open-Sources Aspen Git Server for AI Agent Swarms

13 reports archived9 independent sourcesupdated 11/9/2026, 16:26:22
Synthesis & Latest Updates
9 Sources Cross-Validated

Reason Machines has open-sourced Aspen, a Git server designed for repositories receiving concurrent pushes from hundreds of AI agents. Instead of serializing every push through one repository queue, each branch is a single Postgres row updated via compare-and-swap, so pushes to different branches never block each other. In the team's own test, one 8-core server with S3 Standard sustained 567 pushes per second for five minutes at 208 ms median and 314 ms p99 latency, though landing on main still runs at roughly 4 per second and a 24-hour chaos run stopped at 11.59 clean hours.

LATEST/Reason Machines has open-sourced Aspen, a Git server designed for repositories receiving concurrent pushes from hundreds of AI agents. Instead of serializing every push through one repository queue, each branch is a single Postgres row updated via compare-and-swap, so pushes to different branches never block each other. In the team's own test, one 8-core server with S3 Standard sustained 567 pushes per second for five minutes at 208 ms median and 314 ms p99 latency, though landing on main still runs at roughly 4 per second and a 24-hour chaos run stopped at 11.59 clean hours.

Hourly heat curve

129 fully observed hours
Now
6.1
Peak
122.922:00
24h change
-64%

The line compares only sources observed throughout; gap hours are interpolated to keep the trend continuous.

Coverage timeline

Total 13 reports · Latest first
  2. 新智元 (微信公众号)T2·74 pts

    Tsinghua Open-Sources OpenMise to Give Video Models 3D Spatial Memory

    • Splits generation into 3D scene construction and rendering to enforce cross-frame consistency via explicit spatial state
    • Targets the re-roll pain of frontier models on precise motion and spatial relations, cutting regeneration compute
    • Ships as an agent-callable spatial scene editing tool that can act as a planning layer before video generation
  3. The DecoderT2·58 pts

    Thinking Machines' Inkling 2 Small Model Swears Repeatedly During Benchmark Runs

    • Profanity appeared in 50 of 2,655 runs with 316 total instances, 241 from just two runaway runs
    • One asteroid-trajectory task produced a ~500,000-character repetitive response containing 158 expletives
    • Profanity is not a failure signal: 96% of runs with swearing failed, but overall failure rate is ~75% and 96% of failed runs had no expletives
  5. Arena.ai (@arena)T1·58 pts

    Inkling 2 Small Tops US Open Models on Code Arena WebDev

    • Scored 1654 on Code Arena: WebDev, 128 points ahead of Meta's Muse Spark Open and 245 points above the prior Inkling Small
    • Would rank #1 among US open models, #4 among open models globally, and #20 overall once weights are released
    • Available now only for fine-tuning and inference on Tinker; full weights pending, priced at $0.30/$1.00 per 1M input/output tokens
  6. The DecoderT2·78 pts

    Thinking Machines Releases Inkling 2 Small, Topping Western Open-Weight Models

    • MoE architecture with 594B total and 17B active parameters scores 31 on the Intelligence Index, up 18 points from its predecessor
    • Agentic tasks improve markedly, but cost per task is six times higher with more output tokens
    • Initially proprietary via the Tinker platform, with Apache-licensed weights to follow, more permissive than Meta or Mistral licenses
  8. Marktechpost (@Marktechpost)T2·78 pts

    Thinking Machines Ships Inkling 2 Small, a 594B Sparse MoE Model

    • 594B total parameters with only 17B active per token; sparse routing picks 8 of 384 experts across 48 layers
    • Scores 31 on the AA Intelligence Index and 86.3% on SWE-Bench Pro v2, but only 13.1% on Terminal-Bench, showing uneven capability
    • Available now on Tinker for fine-tuning and inference at $0.30/$1.00 per 1M tokens; weights remain closed for now
  10. Artificial AnalysisT1·72 pts

    Thinking Machines Releases Inkling 2 Small, Scoring 31 on Intelligence Index

    • Scores 31 on the Intelligence Index, up 18 points from Inkling Small and level with Gemini 3.8 Flash (high)
    • Large agentic gains: AutomationBench-AA rises from 5% to 46%, FrontierCode 2.0 from 5% to 25%
    • Coding profile diverges: 25% on FrontierCode 2.0 beats Claude Sonnet 5.5, but 13% on Terminal-Bench 4.0 trails its 30%
    • Currently proprietary with weights pending; if released, it would be the top open-weights model outside China
  12. Artificial Analysis (@ArtificialAnlys)T1·74 pts

    Thinking Machines Releases Inkling 2 Small, Scoring 31 on Intelligence Index

    • Scores 31 on the Intelligence Index, up 18 points from Inkling Small and level with Gemini 3.8 Flash (high), ahead of MiMo-V2.6-Pro
    • Large agentic gains: AutomationBench-AA rises from 5% to 46%, FrontierCode 2.0 from 5% to 25%, and HLE-Diamond from 12% to 29%
    • Coding profile differs from Claude Sonnet 5.5: 25% vs 11% on FrontierCode 2.0, but 13% vs 30% on Terminal-Bench 4.0
    • Priced at $0.30/$1.00 per 1M input/output tokens ($0.03 cached input), available first-party via Tinker at launch