Zhipu AI released GLM-5 on February 14, 2026 — the most significant architectural departure in the GLM series since GLM-4 introduced vision capabilities. GLM-5 moves from GLM-4's dense architecture to a large MoE design at 744B total / 40B active parameters, extends context to 200K tokens, and ships under MIT license — the same open licensing that made DeepSeek V3 attractive for self-hosted deployment.

The timing positions GLM-5 directly against Kimi K2 and DeepSeek V3.2 in the February 2026 crowded-frontier environment. SWE-bench Verified 77.8% places GLM-5 above DeepSeek V3.2 (49.8%) and competitive with Kimi K2 (70.4%) — establishing Zhipu as a serious contender in the agentic coding tier, not just general instruction following.

For developers, GLM-5 is notable as the only large Chinese MoE model with both MIT license and a 200K context window at this capability tier.

What's New

What's New

744B/40B MoE architecture: GLM-5 adopts mixture-of-experts scaling — 744B total parameters with 40B active per token. The MoE design follows the pattern established by DeepSeek V3 and Qwen3: large total capacity at manageable active parameter cost. For the GLM family, this is a significant scale increase from GLM-4's dense architecture.

SWE-bench Verified 77.8%: The headline number — +28pt above DeepSeek V3.2 (49.8%), within 5pt of Kimi K2.6 (80.2%), and competitive with the emerging coding-capable MoE tier. At 77.8%, GLM-5 can handle autonomous code change tasks at quality that was reserved for frontier closed-source models six months prior.

200K token context: GLM-5 extends from GLM-4's 128K to 200K tokens. For full-repository code analysis, legal document review, and multi-document research synthesis, 200K tokens covers most practical use cases without multi-pass chunking.

MIT license: GLM-5 ships under MIT — commercial use, fine-tuning, redistribution, on-premises deployment, no restrictions. This is the most permissive license available for a model at this capability tier, and it distinguishes GLM-5 from Qwen2.5-Max (API-only, proprietary) and Hunyuan-TurboS (proprietary).

Multimodal continuity: GLM-5 retains GLM-4V's vision capabilities — image understanding, chart analysis, and document OCR. The multimodal continuity makes GLM-5 a drop-in upgrade for applications built on GLM-4V, with the added coding and reasoning improvements.

Architecture

Architecture

Component GLM-5 GLM-4 (prev)
Total parameters 744B ~130B dense
Active parameters 40B ~130B
Architecture MoE Dense Transformer
Context window 200K 128K
Vision Yes Yes (4V)
License MIT Restricted
Input price ~¥0.05/K tokens —

MoE routing in GLM-5: GLM-5 uses top-k expert routing with a learned gating mechanism. The MoE transition from GLM-4's dense design represents a fundamental architectural shift — at 744B total vs ~130B active, GLM-5 has roughly 6× the total parameter capacity while maintaining similar inference cost to GLM-4.

Context architecture: 200K context is supported via a combination of RoPE (Rotary Position Embedding) extension and attention optimization. GLM-5 applies techniques similar to YaRN (Yet another RoPE extension) to extend the effective context beyond the training window without full retraining at 200K.

Multimodal integration: GLM-5 maintains the visual grounding capability introduced in GLM-4V — images are encoded into the same token space as text, enabling mixed-modality inputs. Vision capability is particularly relevant for GLM-5's Chinese market focus: OCR of Chinese documents, analysis of Chinese-language charts and infographics.

MIT license implications: Full MIT means: (1) self-host on-premises without licensing fees, (2) fine-tune on proprietary data without restrictions, (3) distribute modified versions without disclosure requirements, (4) use commercially without per-seat or per-call licensing. The only comparable large open MoE is DeepSeek V3 (MIT) — GLM-5 adds vision and a slightly longer context.

Benchmarks

Benchmarks

GLM-5 at February 2026 release:

Benchmark GLM-5 Kimi K2.6 DeepSeek V3.2 Qwen3-72B Notes
SWE-bench Verified 77.8% 80.2% 49.8% 52.1% Real GitHub issues
LiveCodeBench 78.4 81.3 68.9 65.2 Competitive coding
MMLU 88.9% 89.4% 87.8% 87.1% General knowledge
GPQA Diamond 61.2% 63.7% — 57.4% Graduate science
GSM8K 96.1% — 95.7% 94.2% Math word problems
Chinese Bench 92.4% 87.8% 88.1% 91.3% Chinese language

SWE-bench 77.8% is GLM-5's strongest story: +28pt above DeepSeek V3.2, within 2.4pt of Kimi K2.6 (80.2%). For a model launching at MIT license with 200K context, this places GLM-5 in the first tier of open-weight coding models.

Chinese Bench 92.4% is Zhipu's home-market advantage — GLM models have consistently led on Chinese language tasks, and GLM-5 extends this with its multimodal Chinese document processing.

Pricing in Context

Model License Context SWE-bench Input Price
GLM-5 MIT 200K 77.8% ~$0.069/M
DeepSeek V3.2 MIT 256K 49.8% $0.27/M
Kimi K2.6 MIT 256K 80.2% $0.85/M
Qwen3-Coder Proprietary 128K 70.6% $0.048/M

GLM-5 occupies a distinctive position: MIT license + 200K context + SWE-bench 77.8% at ~$0.069/M input. DeepSeek V3.2 (MIT, cheaper) has much lower SWE-bench (49.8%). Kimi K2.6 (MIT, 80.2%) costs more. GLM-5 is the best open-weight option for the specific combination of coding capability + MIT license.

What It Means for Developers

Best open MIT-licensed model for coding + Chinese: With SWE-bench 77.8% and MIT license, GLM-5 is the strongest open-weight option for coding-capable applications that require data residency, on-premises deployment, or fine-tuning. DeepSeek V3.2 (MIT) trails by 28pt on SWE-bench; Kimi K2.6 (MIT) beats by 2.4pt but costs more per API call.

200K context for full-document workflows: Legal contract analysis, full-codebase review, and multi-document research can run in a single context window without chunking. For Chinese-language enterprise customers — a primary GLM deployment context — processing contracts, regulations, and filings in 200K token windows is practical.

Multimodal + text in one model: GLM-5's retained vision capability means applications needing both text and image understanding (document OCR, chart analysis, mixed-content pipelines) use one model rather than routing between a text model and a vision model. For Chinese-language document processing, this is particularly valuable.

Zhipu enterprise relationships: Zhipu AI has deep relationships with Chinese state-owned enterprises and government institutions that require domestic AI vendors. GLM-5 with MIT license and Zhipu's enterprise support structure provides a path to frontier-quality AI for organizations with strict procurement constraints.

Drop-in upgrade from GLM-4V: Same API format, same tokenizer family, same multimodal capabilities — upgraded to 744B MoE, 200K context, and significantly better coding benchmarks. Existing GLM-4V applications upgrade by changing the model identifier.

Availability

Bottom Line

GLM-5 is Zhipu's architectural reset: from dense to 744B MoE, from 128K to 200K context, from restricted to MIT, with SWE-bench 77.8% that places it in the first tier of open coding models. For developers who need MIT license + coding capability + Chinese language strength in a single model, GLM-5 is the clearest option at February 2026. The gap to Kimi K2.6 (2.4pt on SWE-bench, higher API price) leaves GLM-5 as the value option in open-weight frontier coding models — and the dominant choice where Chinese language quality is the primary requirement.

Resources

GLM-5 is available via the Zhipu BigModel Platform at open.bigmodel.cn and on HuggingFace under MIT license.