Inkling 2 Small 594B MoE: 17B Active Parameters

What changed

Thinking Machines Lab released Inkling 2 Small on November 5, the first model in its second generation. The company describes it as a sparse Mixture-of-Experts architecture with 594 billion total parameters and 17 billion active parameters. The model processes text and image input, maintains a one-million-token context window, and supports variable reasoning effort settings. According to the announcement, it scores 31 on the Artificial Analysis Intelligence Index, a jump from the previous Inkling Small’s score of 13 on the same benchmark.

The model launched with closed weights on the Tinker platform for fine-tuning and inference. Thinking Machines plans to release the full weights under an Apache 2.0 license after completing external safety testing and red-teaming, consistent with its stated safe path to open weights policy.

The architecture uses a 48-layer decoder-only transformer where each token routes to 8 of 384 experts. Vision processing relies on a 292-million-parameter encoder based on Perception Encoder, using 20-by-20-pixel patches with 2-by-2 spatial pooling. Image patches are interleaved with text tokens during processing.

Pretraining consisted of four phases across NVIDIA GB300 hardware. The first phase covered 8.4 trillion tokens in 68 hours trained natively in NVFP4 precision. The second phase processed 3.8 trillion tokens of higher-quality data at 512K context. The third phase handled approximately 320 billion tokens of reasoning and agent trajectories at full 1M context. The final phase used approximately 34 billion tokens of highest-quality agent trajectories.

A unified reinforcement learning run followed, taking 1,075 steps and over 17 million rollouts across 2,048 GPUs over 9 days. Compute exceeded 4 times 10 to the 23rd FLOPs. Coding accounted for 65 percent of RL tokens. This single joint run differs from the previous Inkling, which used separate reasoning and agentic RL runs merged through distillation. The company added explicit alignment attention by implementing an alignment reward and on-policy self-distillation from a conduct spec, reducing its internal misbehavior score from 66.8 to 14.9.

Why it matters

The Inkling 2 Small represents a new open-source option in the efficiency-focused MoE landscape. For teams evaluating models, the 17 billion active parameters determine actual inference compute cost, not the full 594 billion parameter count. This sparse routing approach allows developers to run a much larger model without proportional latency or resource overhead. The multimodal capabilities and agentic training expand use cases beyond text-to-text tasks, particularly for coding agents and vision-guided work.

Performance on specific benchmarks reflects capability differences. The model achieves 86.3 percent on SWE-Bench Pro v2 Hard, placing it second only to Kimi K3 in Thinking Machines’ benchmark table. It leads competitors on PerceptionBench with 64.7 percent, EnterpriseBench-CoreCraft at 71.8 percent, and JobBench at 70.5 percent. However, it trails significantly on Terminal-Bench 4.0 with 13.1 percent versus 32.8 percent for GLM-5.3-Flash and 33.8 percent for DeepSeek-V4.1-Flash-1017. The company notes the model runs slowly and verbosely on some task types.

Teams building agentic systems will want to test whether the model’s improvements in instruction following and self-verification translate to fewer hallucinations and better tool-use patterns in production. The announcement emphasizes that Inkling 2 Small follows instructions better, considers changes carefully, and verifies work without being prompted. The alignment improvements, where internal misbehavior scores fell from 66.8 to 14.9, suggest explicit effort to reduce fabrication on hard tasks.

Pricing at $0.30 per million input tokens and $1.00 per million output tokens positions Inkling 2 Small competitively. It costs less on output tokens than the previous Inkling at $1.20 per million, making it attractive for inference-heavy workloads like long agent runs or batch processing.

The model demonstrated cross-domain capabilities in announced tests. On a time-series forecasting task, Inkling 2 Small autonomously fine-tuned Qwen3.5-4B on Tinker, doubling the forecasting score after 4.5 hours of work and $426 in spend. In visual testing, it correctly identified three wines from a single picture in 19 of 20 trials.

What to test

Before adopting Inkling 2 Small in production, teams should verify several claims:

• Agentic reliability: Run the model through your own agent harness and measure error rates, tool-calling accuracy, and unforced termination. Test whether it follows instructions better, verifies work without prompting, and considers changes carefully against your baseline.

• Vision accuracy in your domain: The wine-tasting example is illustrative but domain-specific. If vision is critical, benchmark the model on your own image datasets and compare against GLM-5.3-Flash or Kimi K3 on identical samples.

• Token efficiency under load: Verify that active parameter routing produces the latency and throughput gains your use case requires. Sparse MoE models sometimes need specialized hardware; confirm Tinker’s infrastructure meets your SLA.

• Terminal-Bench weakness impact: Understand whether the 13.1 percent Terminal-Bench 4.0 score affects your use case. Run test queries requiring shell navigation or complex state tracking.

• Alignment robustness in edge cases: Test the reduced misbehavior score by probing with tasks where the model might fabricate results. Does it still confabulate on uncommon technical questions or ambiguous scenarios?

• Weights quality upon release: When Apache 2.0 weights ship, validate that supported quantization formats (MXFP8 and NVFP4) perform as expected for your hardware and inference stack.

• Real-world autoresearch performance: If autonomous research capabilities matter to you, benchmark the model on your own optimization or training tasks to verify the 4.5-hour example generalizes.

The conclusion

Inkling 2 Small 594B MoE offers teams a competitive open-source alternative in the efficiency tier. The 17 billion active parameters and multimodal support align with rising demand for lean agentic models. The 31 Artificial Analysis Intelligence Index score and strong performance on coding and vision benchmarks indicate real capability gains over the previous generation. However, gaps on terminal tasks and the lack of self-hosting details limit early adoption to teams confident in Tinker’s API or willing to wait for weight release and custom deployment.

The planned Apache 2.0 release is significant for reproducibility and fine-tuning at scale. Teams already invested in Thinking Machines’ ecosystem will benefit from improved tool use and reduced hallucination, while those evaluating across vendors should benchmark Inkling 2 Small against similar-scale competitors on your own tasks. The unified RL approach across reasoning and agentic work represents a methodological shift worth studying if you plan to fine-tune for domain-specific agents.

Watch for real-world performance reports after weights drop and external red-teaming concludes. Also track whether self-hosting becomes available and what hardware requirements are published. If latency and throughput hold up in production, this model could become a standard reference point for open MoE inference at this parameter scale.

Related reading

• Prime Mega MoE speeds up mixture-of-experts training 34%
• agent-desktop: Open-source agent computer use
• Gemini Omni 1.2 Flash: 30% price cut, 15-second video