SYLIN

GLM

by Zhipu AI · Model · Language & code models

strong

Solid open model, best for image understanding, no longer the overall leader

#1 image understanding (open weights)

#1 on Arena vision among open weights (1295)

AccessOpen weightsNeeds~756 GBDownloadabout 756 GB (version 5.2 weights)LicenseApache 2.0

Why we say so

Arena ranks GLM 5.3 first among open weights for vision (1295) and second on text, behind MiMo 2.6. Artificial Analysis also placed it above Mistral Large 4 and Haiku 5.5, and Step 5 trails it slightly. MiMo 2.6 overtook it as the top open model on the AA index, DeepSeek's own benchmarks claim V4.1 Flash beats it, and it got only 1 of 6 on AI Search's brain-tumor test.

Versions

VersionReleasedRank
GLM-5.3LATESTAug 14#1 image understanding (open weights)#2 chat (open weights)
GLM-5.2

Evidence

  1. Nov 12
    GLM 5.3 Flash at native precision on 40 Intel N150 mini PCs over USB 3: 14 tok/s at 1k context, 12 at 1M, ~740 W, ~$5,000
  2. Nov 11
    #2 on Arena text among open weights (1509)
  3. Oct 19
    #1 on Arena vision among open weights (1294)
  4. Sep 27
    GLM 5.3 was the top open model on the AA index until MiMo 2.6 Pro; base for Aikido Altar
  5. Sep 20
    GLM 5.3 Flash is multimodal with about 1M context and was deployed on 100,000+ domestic accelerators in under two weeks
Show all 37
  1. Sep 15
    Leading open model; fourth on RealSWE and tied with DeepSeek on Vals Index
  2. Sep 2
    Built a working browser Windows 11 replica in one prompt (about 22 min of thinking), using parallel sub-agents and self-testing; PowerPoint was weak
  3. Aug 30
    GLM-5.3-Flash is the first natively multimodal GLM-5 model: 320B total, 18B active MoE, hybrid sparse and linear attention
  4. Aug 30
    Called the most capable open model available, roughly tied with Kimi K3 despite fewer parameters
  5. Aug 30
    Intelligence index score within one or two points of the best GPT and Claude models
  6. Aug 29
    Fixed a logic bug in a Docker app via Hermes Agent, but took 15 minutes versus Qwen's 4 and tested fractional servings
  7. Aug 29
    Won the 84-language pick-up line test: more varied phrasing, no leaked tags, cleaner low-resource languages like Tigrinya
  8. Aug 28
    Commenters say it feels like Opus 4.8 and handles hard problems better than DeepSeek V4 Flash, though not Fable-class
  9. Aug 28
    Reported as slightly behind Kimi in ability but easier to run, and less restrictive on cybersecurity requests
  10. Aug 28
    Unsloth dynamic 3-bit UD-IQ3 GGUF is 81% smaller than BF16, keeps 82% accuracy, and fits in about 125 GB RAM
  11. Aug 27
    Took #1 on Arena vision among open weights
  12. Aug 26
    Tied for first on the Artificial Analysis agentic index with Claude Opus 5
  13. Aug 26
    In Hermes Agent it found and fixed a safety-separation bug in a multi-container Docker air-traffic app within about 2 minutes, with no hints
  14. Aug 26
    Chart shows it beating GLM-5.2 convincingly and nearing Claude Opus 4.8 on some evals, at a much lower price
  15. Aug 23
    Commenters say security researchers are shifting to Chinese models because US frontier models refuse exploit work
  16. Aug 23
    Ed-o-Meter claims open-weight GLM-5.3 beats Anthropic and OpenAI models at about one fifth the cost
  17. Aug 23
    Gert Labs multi-agent coding evals rank GLM 5.3 around #6; not Pareto optimal versus Grok 4.6 and Sol 5.6
  18. Aug 21
    Commenter switched from Claude to GLM 5.3 over writing style and found it a breath of fresh air
  19. Aug 19
    Scores 59.5 against 53.0 for GLM-5.2, at $0.68 versus $0.56 per task and about 28% more output tokens
  20. Aug 19
    Sits on the cost-intelligence Pareto frontier at about $0.60 per task; Opus and Fable score higher but cost four to five times more
  21. Aug 18
    Similar overall to Kimi K3; better at coding and cybersecurity, Kimi K3 slightly better at visuals, 3D and deep research
  22. Aug 18
    Uses far more output tokens per task than Opus 5 (21,353), GPT-5.6 Sol (16,879) and Kimi K3 (25,474); price-performance edge over K3 is small
  23. Aug 18
    Scores 59.5 on the AA Intelligence Index (max effort) at $0.68 per task and 41,107 output tokens per task
  24. Aug 18
    Scores below Claude Opus 5 (61.5) and GPT-5.6 Sol (60.9) but costs less per task ($0.68 vs $1.52 and $1.23)
  25. Aug 17
    Scored 1 of 6 on brain tumor scan identification despite no vision; only Kimi K3 matched, Opus and Fable scored 0
  26. Aug 17
    Claimed best model on GDPval and Automation Bench, matching Fable and GPT 5.6 on Humanity's Last Exam, number two on Frontier Suite
  27. Aug 17
    Claimed best on CyberGym, ahead of Mythos and GPT 5.6 Sol, and well ahead of Kimi K3 on ExploitBench and ExploitGym
  28. Aug 17
    Said to have found over 2,400 undisclosed vulnerabilities (about 1,000+ high/critical) in Linux kernel, WebKit, FreeBSD; oldest from 1981
  29. Aug 14
    On agentic coding benchmarks incl. Terminal-Bench, Agents' Last Exam, AutomationBench, GDPval it sits at or near the top of open models, near closed frontier
  30. Aug 14
    Big jump over GLM 5.2 on white-box vulnerability finding, exploit reasoning and timed exploit challenges; top of open models, sometimes edging closed ones
  31. Jul 26
    Hugging Face reportedly used open-weights GLM-5.2 to detect and fix the breach because closed models refused the task
  32. Jul 17
    On a one-file concrete-plant simulation prompt, GLM 5.2 gave a working dark-theme UI with passing self-checks; judged third of three

Cost to run

Alternatives

Kimi#3 on Arena text among open weights (1504)
Claude OpusGot 1 of 6 brain-tumor types right, tied with Astra, GLM 5.3 and Chimera 3 as the best so far
Claude FableAutonomously researched earnings and produced a motion-graphics financial video; layout looked a bit better than GLM 5.3 and GPT output
MiMo#1 on Arena text among open weights (1534)
DeepSeek FlashBeats GLM 5.3 and Kimi K3 across the shown benchmarks, including agentic coding and general knowledge
More GLM