Reinforcement.tech — Models
Tinker catalog

Models

Thinking MachinesTinker

The compiler writes one BASE_MODEL string. Inkling and Inkling-Small are Tinker’s own weights — they compile. Interaction models are a real-time research preview, not a hosted base_model, so they are not on this desk. Full hosted list: Tinker models.

20 base models, curated. Tinker’s full hosted list is on Models & Pricing. This desk writes those IDs; Tinker may host more.

  • thinkingmachines/Inkling
  • thinkingmachines/Inkling-Small
  • Qwen/Qwen3.5-4B
  • Qwen/Qwen3-8B
  • Qwen/Qwen3.5-9B
  • Qwen/Qwen3.8-27B
  • nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-BF16
  • nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16
  • Qwen/Qwen3.6-35B-A3B
  • openai/gpt-oss-20b
  • openai/gpt-oss-120b
  • zai-org/GLM-5.3:peft:262144
  • moonshotai/Kimi-K2.6
  • deepseek-ai/DeepSeek-V3.1
  • nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B-BF16
  • nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-BF16
  • Qwen/Qwen3.6-27B
  • Qwen/Qwen3.5-397B-A17B
  • Qwen/Qwen3.5-35B-A3B-Base
  • Qwen/Qwen3.5-9B-Base

Thinking Machines

Inklingthinkingmachines/Inkling · Hybrid + Audio + Vision · MOE · 975B-A41B · 64KThinking Machines’ own weights. Hybrid + audio + vision.For: Calibrated Forecasting · Formal Reasoning EnginesSample $4.68/M tok · train $5.61/M tokIntroducing InklingBuild
Inkling-Smallthinkingmachines/Inkling-Small · Hybrid + Audio + Vision · MOE · 276B-A12B · 64KTinker’s smaller Inkling. Latency-friendlier own-weights pick.For: Specialized Tool AgentsSample $1.44/M tok · train $1.73/M tokIntroducing Inkling-SmallBuild

Dense — every weight live

Qwen3.5-4BQwen/Qwen3.5-4B · Hybrid + Vision · DENSE · 4B dense · 64KSmallest dense base model before you scale rank or data.For: any jobSample $1.005/M tok · train $0.737/M tokBuild
Qwen3-8BQwen/Qwen3-8B · Hybrid · DENSE · 8B dense · 32KDense workhorse when you want every weight live in the loop.For: any jobSample $0.60/M tok · train $0.44/M tokBuild
Qwen3.5-9BQwen/Qwen3.5-9B · Hybrid + Vision · DENSE · 9B dense + vision · 64KWhen the reward needs screenshots, diagrams, or Lean goals.For: any jobSample $1.995/M tok · train $1.463/M tokBuild
Qwen3.8-27BQwen/Qwen3.8-27B · Hybrid + Vision · DENSE · 27B dense + vision · 64KCurrent dense mid-size on Tinker. Successor weight class to Qwen3.6-27B.For: any jobSample $5.595/M tok · train $4.103/M tokBuild
Qwen3.6-27BQwen/Qwen3.6-27B · Hybrid + Vision · DENSE · 27B dense + vision · 64KStill listed on Tinker. Prefer Qwen3.8-27B for new work.For: any jobSample $5.595/M tok · train $4.103/M tokBuild

MoE — sparse-activate

Nemotron-3.5-Lightningnvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-BF16 · Hybrid · MOE · 30B-A3B · 64KLatency-first MoE. The current Lightning pick next to Nano.For: Specialized Tool AgentsSample $0.495/M tok · train $0.44/M tokBuild
Nemotron-3-Nanonvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16 · Hybrid · MOE · 30B-A3B · 64KCheap active params, fast tool planning.For: Specialized Tool AgentsSample $0.495/M tok · train $0.44/M tokBuild
Qwen3.6-35B-A3BQwen/Qwen3.6-35B-A3B · Hybrid + Vision · MOE · 35B-A3B + vision · 64KDefault Tinker mid-size: thrifty experts.For: any jobSample $1.335/M tok · train $1.177/M tokBuild
GPT-OSS-20Bopenai/gpt-oss-20b · Reasoning · MOE · 21B-A3.6B · 32KOpen-weight MoE when you want GPT-OSS without the 120B bill.For: any jobSample $0.45/M tok · train $0.396/M tokBuild
GPT-OSS-120Bopenai/gpt-oss-120b · Reasoning · MOE · 117B-A5B · 32KThe GPT-OSS Tinker hosts for mid-size reasoning adapters.For: Calibrated ForecastingSample $0.84/M tok · train $0.737/M tokBuild
GLM-5.3zai-org/GLM-5.3:peft:262144 · Reasoning · MOE · 753B-A40B · 256KZ.AI on Tinker. Always-on chain of thought.For: Formal Reasoning EnginesSample $12.15/M tok · train $14.58/M tokBuild
Kimi-K2.6moonshotai/Kimi-K2.6 · Hybrid + Vision · MOE · 1T-A32B · 32KThe large-end model Tinker’s own FAQ names.For: Calibrated ForecastingSample $5.49/M tok · train $4.84/M tokBuild
DeepSeek-V3.1deepseek-ai/DeepSeek-V3.1 · Hybrid · MOE · large MoE · 32KHard jobs: small adapter on a huge base model.For: Formal Reasoning EnginesSample $4.215/M tok · train $3.718/M tokBuild
Nemotron-3-Ultranvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B-BF16 · Hybrid · MOE · 550B-A55B · 64KWhen the job wants Nemotron scale without leaving Tinker.For: any jobSample $6.225/M tok · train $5.478/M tokBuild
Nemotron-3-Supernvidia/NVIDIA-Nemotron-3-Super-120B-A12B-BF16 · Hybrid · MOE · 120B-A12B · 64KBetween Lightning and Ultra when Nano is too small.For: any jobSample $1.44/M tok · train $1.276/M tokBuild
Qwen3.5-397B-A17BQwen/Qwen3.5-397B-A17B · Hybrid + Vision · MOE · 397B-A17B + vision · 64KWhen the adapter should sit on a wide Qwen expert mix.For: any jobSample $7.50/M tok · train $6.60/M tokBuild

Base weights — no instruct tuning

Raw pretrained weights. The usual starting point for RL research.

Qwen3.5-35B-A3B-BaseQwen/Qwen3.5-35B-A3B-Base · Base · MOE · 35B-A3B base · 64KRL researchers start here: no chat prior to fight.For: any jobSample $1.335/M tok · train $1.177/M tokBuild
Qwen3.5-9B-BaseQwen/Qwen3.5-9B-Base · Base · DENSE · 9B dense base · 64KSmaller base-not-instruct Qwen for recipe search.For: any jobSample $1.995/M tok · train $1.463/M tokBuild
Models — Reinforcement: Build Your Own Reward Model