Inkling 2 Small reportedly beats GLM-5.3 and DeepSeek V4 on code with 40% of their pretraining compute
teortaxesTex · x · 2026-11-06
A leaked training-cost comparison shows Inkling 2 Small using only 1.3×10²⁴ FLOPs of pretraining compute (about 40% of DeepSeek V4.1 Flash's), with 17B active params × 12.6T tokens, yet reaching 0.2030 bits-per-byte on Internal Code — better than GLM-5.3 (0.2358), DeepSeek V4 Pro Base (0.2335), and Kimi K3.1 Pro (0.2068).
Training ran four phases (8.4T general → 3.8T high-quality → reasoning/agent trajectories → top-quality agent data) plus 9 days of RL, 13 days total and 1.7×10²⁴ FLOPs including RL. The poster argues that if the official numbers fold RL into "pretraining," the advantage over pretraining-only rivals like DeepSeek is understated — a strong base trained cheaply, with RL coming easy afterward.
Related event: Thinking Machines Releases Inkling 2 Small Open-Weight Model(42 posts)→
More from Infra
- Zyphra ships ZYFA: FP8 attention on AMD MI355X at 1.16-1.76x AITER throughput — bclavie · 2026-11-06
- Akamai's signed AI book tops $14B, anchored by Anthropic's $11.6B commitment — pdamodaran · 2026-11-06
- Teknium pushes back on METR's 'no bottlenecks' claim, citing funding and compute cost inflation — Teknium · 2026-11-06
- Fudan's D3D decouples diffusion drafting for asynchronous speculative decoding speedups — Fudan-University · 2026-11-06
- ThursdAI pod: midterms, 40 datacenter ballot measures, $3T wipeout — is AI over? — altryne · 2026-11-06
- sshhole Bug Analysis: Alpine/MUSL Systems Hit Hard, ~100-500M OpenWRT Routers Were at Risk — esrtweet · 2026-11-06