Inkling 2 Small beats Kimi K3.1 on code loss with just 12T pretrain tokens
teortaxesTex · x · 2026-11-06
teortaxesTex calls the newly released open-weight Inkling 2 Small (17B active / 594B total params) one of the most impressive Western open-weight releases this year:
- Pretrained from scratch in only a few dozen days on just 12T tokens, far less compute and data than competitors
- Base-model ppl loss on internal code beats Kimi K3.1, even after K3.1 was "base-recovered" for a fair comparison
- Strong at vision and modern frontier coding tasks, and cheap — RL on the strong base was easy
Screenshots show: on internal code eval, Inkling 2 Small (0.2030 pre-trained / 0.2080 post-trained bits per byte) outperforms base-recovered GLM-5.3 and DeepSeek V4 Pro Base; a single joint RL run took 9 days on 2,048 GPUs with 1,075 steps and 17M+ rollouts (>4×10²³ FLOPs), replacing the previous separate reasoning/agentic RL plus distillation approach; public benchmarks lead GLM-5.3-Flash, MiniMax-M3.1-Flash, and DeepSeek-V4.1-Flash across design, visual perception, and STEM reasoning categories.
Related event: Thinking Machines Releases Inkling 2 Small Open-Weight Model(42 posts)→
More from Models
- AI Cyber Skill Doubles Every 4-5 Months, Task Horizon Grows 52x in a Year — FinanceYF5 · 2026-11-06
- 50 insiders vote on AI data companies shaping frontier research: Surge leads with 16 votes — FinanceYF5 · 2026-11-06
- OpenAI publishes two more misalignment incident reports: grader prompt injection and sandbox network bypass — scaling01 · 2026-11-06
- Thinking Machines releases Inkling 2 Small; RL training taught model to circumvent guardrails — johnschulman2 · 2026-11-06
- Qwen3.8-Flash-Next FP4 quantized model trends on Hugging Face — autotrust · 2026-11-06
- Small model Inkling 2 Small ranks #2 on SWE Atlas, matches Kimi-K3 on SWE-Bench Pro V2 — simonguozirui · 2026-11-06