Inkling-Small: 276B params, 12B active, open-source tops ARC

Thinking Machines Lab released its second open-source model, Inkling-Small, on July 31. It is a MoE model with 276B total parameters and only 12B active, supporting a 1M token context and image/audio inputs, with weights fully open under Apache 2.0. Community evaluations show it matches or exceeds the original Inkling with a fraction of active parameters, excelling in coding, agentic tasks, and reasoning, and it became the highest-scoring open-weight model on ARC-AGI-1 and ARC-AGI-2, with a single inference cost of $0.23. Deployment is extremely accessible, sparking widespread community interest.

Confirmed

• Architecture and specs: MoE, 276B total, 12B active, 1M context, variable thinking effort, natively multimodal (image/audio input and conversational interaction). Trained on Nvidia GB300 NVL72.
• Openness and performance: Weights fully open under Apache 2.0. @ArtificialAnlys noted it rivals flagship models with only one-third the active parameters; @simonguozirui relayed that it surpasses same-tier models and the larger original Inkling in agentic tasks and reasoning; @mervenoyann confirmed coding ability exceeds the larger version; @rohanpaulai added HLE score of 31.6% vs Inkling's 29.7%; @keirp1 revealed it is the highest-scoring open-weight model on ARC-AGI-1 and ARC-AGI-2 in ARC Prize official validation, with $0.23 per inference.
• Deployment and quantization: @pcuenq confirmed nvfp4 checkpoint runs on a single Mac Studio (MLX) (150GB), even across two 128GB laptops. @woosukk emphasized it fits on a single NVIDIA B300. @danielhanchen noted Unsloth provided dynamic quantization; @baseten confirmed availability on its platform.

Why it matters

• Cost-efficiency and low barrier: Inkling-Small demonstrates extreme parameter efficiency, achieving top-tier performance with minimal active parameters, challenging the notion that large models require massive compute clusters. This enables developers to run a 100B-class model on local hardware, drastically lowering the cost and barrier to experimenting with frontier AI.

2026-07-30 ~ 2026-08-01 · 48 related posts

Full story(2 episodes)→

• Episode 1: Inkling-Small: 276B params, 12B active, open-source tops ARC(2026-07-30, 48 posts)
• Episode 2: Inkling-Small Tops ARC-AGI Open-Source Leaderboard at Record Low Cost(2026-07-31, 4 posts)

• Unsloth Releases Inkling-Small Multimodal MoE Model — unsloth · 2026-07-30
• Inkling-Small Released: 276B Parameter MoE Model Matches Original Performance — ziqiao_ma · 2026-07-31
• Thinking Machines Releases Inkling-Small: 276B Parameters Matching Original Performance — simonguozirui · 2026-07-31
• Thinking Machines Releases Inkling-Small Model — NielsRogge · 2026-07-31
• Inkling Small launches: 276B parameter model runs on a single Mac Studio with MLX — pcuenq · 2026-07-31
• [source] thinkingmachines Releases Inkling-Small: 276B Params, 1M Context — rerri · 2026-07-31
• Thinking Machines Launches Inkling-Small: 12B Active Params Beats Larger Model — mervenoyann · 2026-07-31
• Thinking Machines Launches Inkling Small: Open-Weights Model Matches Flagship at 1/3 Size — ArtificialAnlys · 2026-07-31
• Thinking Machines Launches Inkling Small: 12B Active Params Matches Flagship Performance — ArtificialAnlys · 2026-07-31
• Inkling-Small Released: Matches Original Performance at 1/4 the Size — arshdeep · 2026-07-31
• 276B Param MoE Model Runs on Single B300, vLLM Creator Praises Inkling-Small — woosuk_k · 2026-07-31
• Running 276B Inkling Small on a Single Mac Studio via nvfp4 Quantization — pcuenq · 2026-07-31
• Inkling-Small Released: 276B Parameter MoE Model — simonguozirui · 2026-07-31
• Tinky Machines Releases Inkling-Small: 276B Parameters with Open Weights — simonguozirui · 2026-07-31
• Inkling-Small Released: Beats Larger Sibling at 1/4 the Size — simonguozirui · 2026-07-31
• Thinking Machines Releases 276B Open Model Inkling-Small — danielhanchen · 2026-07-31
• Thinking Machines Launches 276B Inkling-Small for Ultra-Fast Speech-to-Speech — MaziyarPanahi · 2026-07-31
• Inkling-Small Released: Matches Larger Model at 1/4 Size via Distillation — simonguozirui · 2026-07-31
• Inkling-Small Released: 12B Active Params Delivers Comparable Performance — clarejtbirch · 2026-07-31
• [source] Inkling-Small: New MoE Model for Image/Audio-to-Text Trends on Hugging Face — thinkingmachines · 2026-07-31
• Inkling Small Tops ARC-AGI Open-Weight Leaderboard at $0.23/Task — keirp1 · 2026-07-31
• Thinking Machines Releases Inkling-Small: A 276B MoE Open Weight Model — LiTianleli · 2026-07-31
• Inkling-Small (12B Parameters) Ranks Alongside 2-3x Larger Open Models — simonguozirui · 2026-07-31
• Mira Murati's startup releases open-source Inkling-Small model — simonguozirui · 2026-07-31
• LMSYS Releases Inkling-small: 12B Active Parameters Lowers Multimodal RL Barrier — simonguozirui · 2026-07-31
• Thinkymachines releases Inkling-Small: 276B-parameter open-source multimodal model with 140 tps inference at $11 — _akhaliq · 2026-07-31
• Thinking Machines Launches Inkling-Small: A 276B MoE Model with Variable Thinking Effort — clarejtbirch · 2026-07-31
• Thinking Machines Launches Inkling-Small: 12B Active Params Beats Trillion-Param Model — 新智元 · 2026-07-31
• Inkling-Small Launches: 276B Params, 12B Active at One-Fifth the Cost — togethercompute · 2026-07-31
• Inkling-Small Available on Together AI for Cost-Effective Agentic Coding — togethercompute · 2026-07-31
• ThinkingMachines' Distilled Small Model Beats Large Model in Coding and Reasoning — 量子位 · 2026-07-31
• Thinking Machines releases open-weight Inkling-Small: 12B active params, beats larger Inkling on HLE (31.6% vs 29.7%) — rohanpaul_ai · 2026-07-31
• Thinking Machines Launches 276B Inkling-Small, Rivaling Larger Models with 12B Active Params — arena · 2026-08-01
• Thinking Machines Launches Inkling Small, Beating Full-Size Model in Benchmarks — soumitrashukla9 · 2026-08-01
• Inkling-Small Launches on Together AI: 276B MoE with 1M Context — togethercompute · 2026-08-01
• Thinking Machines Launches Inkling Small: Quarter the Size, Beats Original on Vals Index — ziqiao_ma · 2026-08-01

12 near-duplicate retellings: simonguozirui · simonguozirui · ArtificialAnlys · simonguozirui · baseten · LiTianleli · clarejtbirch · gaganghotra_ · simonguozirui · zdhnarsil · arena · simonguozirui