SYLIN

Ling

by inclusionAI · Model · Language & code models

promising

Capable image-reading and coding model, but thinks slowly and misreads details

Inclusion AI released Ling 3.2 Flash, a hybrid reasoning model in the Ling family

AccessOpen weightsPrice$0 in · $0 out per M tokensContext1,049K tokensNeeds~574 GBDownloadabout 574 GB for 3.2 Flash

Why we say so

Fahd Mirza built a working app from 8 images and extracted an invoice, but the build took 1h19m and he saw a misread year and weak variety in rare languages. He sees 3.2 as a gradual step up from the previous Ling, and advises turning thinking down. The top-spot chart claims on documents and screenshot-to-code come from Inclusion AI, which also concedes it trails GPT and Claude on the hardest multimodal exam.

Versions

VersionReleasedRank
Ling 3.2LATEST
promising Better than before in vision and coding, but slow when thinking
Nov 8
Ling 3.1
unproven Only the maker's claim of being close to GLM 5.3 and Kimi K3
Oct 18
Ling 3.0
unproven No evidence in this record
Jul 23

Evidence

  1. Nov 17
    Ling 3.2 Flash: multimodal 560B MoE, 25B active, text, image and video input; API only, weights not yet released
  2. Nov 17
    Extracted a photographed invoice to JSON, verified arithmetic, read stamp and signature; misread invoice year 2026 as 2016
  3. Nov 17
    Chart shows top spot on document understanding against GPT Sol, Claude Opus, GLM 5.3 Flash and DeepSeek V4 Flash at max settings
  4. Nov 17
    Chart shows top spot on turning a website screenshot into working code; also top on medical answer correctness
  5. Nov 17
    Trails GPT and Claude on the hardest multimodal exam and a few agentic tasks, per the vendor chart
Show all 9
  1. Oct 18
    Ling 3.1 Flash: 560B MoE (25B active), about 574 GB, open weights; close to GLM 5.3, Kimi K3 and DeepSeek 4.1 on benchmarks
  2. Jul 23
    Ling 3.0 Flash, a 124B MoE with 5.1B active parameters, offered via free OpenRouter API; weights not on Hugging Face
  3. Jul 23
    Claimed to match its own 1T flagship on most benchmarks with 1/8 the total and 1/12 the active parameters
  4. Jul 23
    Benchmarks claim parity or leads on agentic tasks, multilingual coding and long-context retrieval against models with 10-20x more active parameters

Cost to run