Thinking Machines Ships Inkling 2 Small: Beats Mistral in Open Weights

(SAN FRANCISCO —) Europe’s brief hegemony over Western open-source artificial intelligence has officially drawn to a close. Thinking Machines, the high-profile start-up founded by former OpenAI Chief Technology Officer Mira Murati, has pulled the curtain back on Inkling 2 Small. Verified by benchmarking service Artificial Analysis, the new architecture leaps ahead of Mistral Large 4 to claim the crown as the most powerful open-weight model engineered outside of mainland China. This engineering milestone arrives concurrently with intense financial maneuvers, as Murati reportedly spearheads a massive $5 billion funding round targeting a $40 billion pre-money valuation backed by venture firm Accel.

The release underscores a fierce geopolitical and corporate race for open frontier supremacy. While Chinese models like DeepSeek V4.1 Flash, Z.ai GLM-5.3, and Moonshot AI's Kimi K3.1 Pro continue to dominate the top tier of open benchmarks, Inkling 2 Small establishes a formidable Western beachhead. However, a granular examination of Artificial Analysis telemetry reveals a complex computational trade-off: extraordinary agentic reasoning gains accompanied by ballooning token consumption, higher economic costs, and deliberate behavioral constraints in query handling.

Technical Architecture & Engineering Breakdown

Inkling 2 Small represents Thinking Machines' second-generation leap in scaling transformer efficiencies for autonomous software agents. Unlike standard instruction-tuned checkpoints designed for single-turn dialogue, the underlying architecture is fine-tuned explicitly for multi-step reasoning traces. Telemetry indicates that the model excels when deployed in autonomous loops where an AI agent must self-direct over prolonged temporal horizons—submitting API calls, parsing execution errors, and dynamically refactoring execution paths.

Yet, these capabilities demand heavy computational overhead. Inkling 2 Small processes approximately 150,000 output tokens per index benchmark task. This represents a fourfold expansion over its predecessor and nearly five times the token burn rate of Anthropic’s Claude Sonnet 5.5. This expansive chain-of-thought generation allows the system to dissect complex engineering challenges, but it directly impacts operational expenditure and inference latency across clusters.

Benchmark Evaluation & Rigorous Comparison

On Artificial Analysis's consolidated evaluation grid, Inkling 2 Small logs a score of 26, matching Mistral Large 4 and comfortably outperforming Meta’s Muse Spark Open at 25. Nevertheless, the global open-weights landscape remains fiercely contested, with Chinese laboratories maintaining a decisive edge. Models such as DeepSeek V4.1 Flash and Z.ai’s GLM-5.3 hit 32 points, while Moonshot AI's newly released Kimi K3.1 Pro weights on Hugging Face achieve an imposing 46 points.

A closer inspection of model telemetry by independent evaluators highlights a nuanced nuance behind the model's low hallucination rate. The reduction in erroneous outputs is partly an artifact of refusal architecture: Inkling 2 Small's answer attempt rate drops from 77 percent down to 53 percent when confronting ambiguous or highly complex prompts. Consequently, while raw accuracy on attempted queries edges up from 32 percent to 34 percent, the model deliberately bypasses a larger fraction of edge-case inputs to preserve statistical precision.

Deployment, VRAM Footprint & Inference Optimization

From an infrastructure standpoint, running Inkling 2 Small locally demands specialized hardware configurations. While exact quantization guidelines and vLLM parameter matrices are pending the official weights drop, enterprise deployments will require dense VRAM allocations to handle its long-context chain-of-thought generation. Each task execution costs approximately $1.18. While this is nearly six times more expensive than the original Inkling Small, it remains highly cost-effective compared to proprietary giants like Claude Sonnet 5.5 ($10.02 per task) and Gemini 3.8 Flash ($7.77 per task).

Thinking Machines' corporate trajectory is similarly high-octane. Following a bumpy period marked by co-founder departures back to OpenAI and Meta, the start-up is securing formidable infrastructure partnerships. NVIDIA remains firmly onboard, slated to supply Thinking Machines with at least one gigawatt of compute capacity derived from next-generation silicon architectures. As negotiations for the $5 billion Accel-led financing round progress, the pressure on Mira Murati’s team to transition Inkling 2 Small from paper to accessible open weights has never been higher.

# Clone weights repository upon official release
git clone https://huggingface.co/thinking-machines/inkling-2-small

# Launch vLLM optimized inference server with long-context support
python -m vllm.entrypoints.openai.api_server \
    --model thinking-machines/inkling-2-small \
    --tensor-parallel-size 4 \
    --max-model-len 131072 \
    --gpu-memory-utilization 0.92