Inkling-2-Small completes 911 enterprise tasks/hour on one B300 node, 43% cheaper than DeepSeek
ryanshrout · x · 2026-11-17
Signal65 benchmarked Inkling-2-Small, the new open-weight model from Thinking Machines Lab, on enterprise agentic workloads. It scores 4,101 on the PINNACLE Model Score — the cheapest correct task among open-weight models above 4,000.
Key numbers:
- One 8×NVIDIA B300 node completed 911 correct multi-step enterprise tasks per hour, 1.8x the next open-weight model above 4,000
- 48 concurrent agents each sustained 96 tok/s, versus roughly 10–15 tok/s per agent for comparable open models under load
- At $76/hr for a rented node, a correct task costs $0.08 — 43% cheaper than DeepSeek-V4.1-Flash-1017 and 59% cheaper than Qwen3.8-Flash-Next
- 93.9% job completion, 97.0% retrieval fidelity; its 4.9% fabrication rate is the highest among open models above 4,000
The pitch: self-hosting enterprises no longer have to choose between a high-scoring model and a fast one.
More from Infra
- Researchers Find Undocumented x86 Instructions Allowing User-Space Microcode Writes — sull · 2026-11-17
- Blackstone data center SPV raises $300M of $800M target as debt pipeline chokes — edgarpavlovsky · 2026-11-17
- MS outage: latent bug from 4-month-old commit, human-written code, AI helped diagnosis — sebkrier · 2026-11-17
- Inside the training-data business: $2k-5k per problem, batches priced in millions — teortaxesTex · 2026-11-17
- Nvidia's problem: OpenAI is "90 plus percent Nvidia" and burning the most cash — ThereWas · 2026-11-17
- Branchy: branched training makes multi-device distributed LLM inference 1.5-2.6x faster — ManuXD32 · 2026-11-17