Inkling-Small Release: The New Open-Source Champion

Episode 1 · Inkling Tops ARC-AGI Leaderboard for Open-Weight Models (2026-07-17, 2 posts)

The open-weight model Inkling has achieved the highest scores to date on both the ARC-AGI-1 and ARC-AGI-2 benchmarks. Notably, it reached a 36.5% accuracy on ARC-AGI-2 with a low cost of $0.64 per task.

• Inkling Sets New Record on ARC-AGI — shizhediao · 2026-07-17
• Inkling Tops ARC Open-Weight Leaderboard — GregKamradt · 2026-07-18

Episode 2 · Thinking Machines Releases Inkling-Small: 276B MoE Open-Source Model (2026-07-30, 31 posts)

Founded by former OpenAI CTO Mira Murati, Thinking Machines has officially released its second open-source model, Inkling-Small. Utilizing a MoE architecture with 276B total parameters and only 12B active parameters, it matches the performance of the original Inkling at a quarter of the size. It also becomes the top-scoring open-weight model on the ARC-AGI benchmark. Supporting a 1 million token context and native multimodal capabilities, the model significantly lowers deployment barriers, drawing massive community attention.

已确认

• 模型架构与规格: Inkling-Small adopts a Mixture-of-Experts (MoE) architecture with 276B total parameters and 12B active parameters per token. It supports a 1 million token context window, features variable thinking intensity, and handles native multimodal inputs from image and audio to text.
• 开源与性能表现: The model's weights are fully open under the Apache 2.0 license. Performance-wise, @ArtificialAnlys notes it rivals flagship models with only a third of the active parameters; @simonguozirui shared that it surpasses peer models and even larger versions in agentic and reasoning tasks; @keirp1 revealed that in official ARC Prize validation, it became the highest-scoring open-weight model on the ARC-AGI-1 and ARC-AGI-2 benchmarks, with a single inference cost of $0.23.
• 部署与量化支持: Various deployment options are available from the official team and community. @pcuenq confirmed its nvfp4 checkpoint runs on a single Mac Studio (MLX) using about 150GB, and can even operate across two 128GB laptops. @woosukk highlighted that the model is small enough to run on a single NVIDIA B300 accelerator. Additionally, the Unsloth team provided dynamic quantization, and the Baseten platform has simultaneously launched the model.

为什么重要

• 高性价比与低门槛: Inkling-Small demonstrates extreme parameter efficiency, achieving top-tier performance with very few active parameters and breaking the stereotype that massive models must rely on huge compute clusters. This allows developers to locally run 100B+ parameter models on consumer-grade or light enterprise hardware, greatly lowering the barrier to trialing and deploying cutting-edge AI models.

• 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
• Thinking Machines Releases Inkling-Small: 276B Parameters Matching Original Performance — simonguozirui · 2026-07-31
• Inkling Small launches: 276B parameter model runs on a single Mac Studio with MLX — pcuenq · 2026-07-31
• 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: 276B Params, Only 12B Active — simonguozirui · 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
• Mira Murati's Inkling Small Matches Flagship at 1/3 Parameters — ArtificialAnlys · 2026-07-31
• Running 276B Inkling Small on a Single Mac Studio via nvfp4 Quantization — pcuenq · 2026-07-31
• Mira Murati's Startup Launches Inkling Small Model — simonguozirui · 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

11 more related posts →

Episode 3 · Inkling-Small Tops AudioMC and Sets Open-Source Record on ARC-AGI (2026-07-31, 2 posts)

Thinky Machines' Inkling-Small model showcases exceptional performance by topping the AudioMC leaderboard for audio capabilities and securing second place in tool calling, just behind Kimi K3. It also achieved a record-breaking 40.1% score on ARC-AGI-2 for open-weights models.

• Inkling-Small Sets New Open-Weight Record on ARC-AGI Benchmark — tessybarton · 2026-07-31
• Inkling-Small Ties for 1st on AudioMC, Ranks 2nd in Open Tool Calling — ziqiao_ma · 2026-07-31