SYLIN

Inkling-Small

by Thinking Machines Lab · Model · Language & code models

promising

Strong open model for coding and agent work, but large and newly released

Inkling 2 Small open weights are now on the Hugging Face Hub

AccessOpen weightsPrice$0.3 in · $1 out per M tokensContext1,049K tokensNeeds~532 GB; smallest build ~160 GBDownloadabout 393 GB; 4-bit quants need about 160 GBLicenseApache 2.0

Why we say so

Arena puts Inkling-2-Small second on agent tasks and third on web development among open weights, behind Kimi K3.1, and MathArena has it top on BrokenArXiv. SpeedRunBench has it clearing a game level ahead of Opus 5 and GPT-5.6 Sol, a single narrow test. The claim that it cheats less than most models comes from its own team, and there is no hands-on review yet.

Ranks #2 for agents (open weights) · #3 for web development (open weights)

Evidence

  1. Nov 17
    Cleared SuperTux 1-1 in 1,988 frames on SpeedRunBench, ahead of Opus 5 and GPT-5.6 Sol
  2. Nov 15
    #2 on Arena agent among open weights
  3. Nov 15
    #3 on Arena webdev among open weights (1644)
  4. Nov 14
    In one experiment it independently trained and improved a time-series forecasting model, designing experiments and adjusting training
  5. Nov 14
    #1 US open model on Code Arena WebDev at 1647; #2 open on Agent Arena at $0.26 per task; top open on MathArena BrokenArXiv
Show all 14
  1. Nov 13
    594B MoE, 17B active, strong at coding, vision and long autonomous tasks; on Tinker now, Apache 2.0 weights promised next week
  2. Nov 11
    #3 on Arena webdev among open weights (1647)
  3. Aug 2
    Inkling Small: omnimodal (text, audio, image, video), 276B total with 12B active, about a quarter of full Inkling; 532 GB
  4. Aug 2
    Holds up against DeepSeek V4 Flash (old version) and GPT-6 Luna; its edge is audio input, which those lack
  5. Aug 2
    One point below full Inkling on Artificial Analysis, but behind the latest DeepSeek V4 Flash in intelligence
  6. Jul 30
    Converted a form image to structured JSON quickly and without missed fields; vision and audio judged good quality
  7. Jul 30
    Claimed to match or beat the larger Inkling on reasoning and agentic coding tasks, with 12B vs 41B active parameters
  8. Jul 19
    Open-weights multimodal MoE: ~975B total, 41B active, 1M context, trained on 45T tokens of text, image, audio and video
  9. Jul 19
    Trails GLM 5.2 and Kimi K3 on Artificial Analysis and is far slower than MiniMax M3 and GLM 5.2; strength is audio and vision

Cost to run

Alternatives

Kimi#1 on Arena agent among open weights
DeepSeek Flash#2 on Arena webdev among open weights (1659)
MiMo#3 on Arena agent among open weights
Made for Inkling-Small