GLM-5.1: 754B Open-Weight Agentic Coding Model, 200K Context, MIT

TL;DR

Last updated July 2026.

GLM-5.1 is an open-weight (MIT) mixture-of-experts coding model from Z.ai (formerly Zhipu AI), released April 7, 2026. It has 754B total parameters with 40B active per token, a 200K-token context window, and 128K max output. At launch it took the top SWE-bench Pro score at 58.4, ahead of GPT-5.4 (57.7) and Claude Opus 4.6 (57.3), and scored 63.5 on Terminal-Bench 2.0. Z.ai list pricing is $1.40/M input and $4.40/M output, and cheaper third-party hosts start near $1.05/$3.50.

What Is GLM-5.1?

GLM-5.1 is an open-weight mixture-of-experts coding model from Z.ai (formerly Zhipu AI), the lab behind the GLM / ChatGLM lineage, released April 7, 2026 under the MIT license. It has 754B total parameters (40B active per token), a 200K-token context window, and up to 128K-131K max output. It shipped as a post-training upgrade to GLM-5, with a roughly 28% coding improvement over the base model from refined post-training and asynchronous agent RL. The weights are on Hugging Face (zai-org/GLM-5.1) under MIT.

It is a coding-first, agent-first model. Z.ai reports it sustains autonomous execution for up to 8 hours across hundreds of tool-call rounds, and it took the top SWE-bench Pro score at launch. It runs on transformers, vLLM, SGLang, KTransformers, and xLLM, and drops into Claude Code or Cline with a config change. Its successor, GLM-5.2 (753B, 1M context), shipped in June 2026; Morph serves that newer model as morph-glm52-744b on custom kernels via an OpenAI-compatible API. See GLM-5.2, Morph Open Source Models, and pricing.

Architecture: DSA Sparse Attention

GLM-5.1 is a 754B-total mixture-of-experts model with 40B active per token, built on the GLM-5 architecture. Its long-context efficiency comes from DeepSeek Sparse Attention (DSA) layered on Multi-head Latent Attention (MLA): rather than every token attending to every previous token, the current token attends only to itself and a selected subset of prior tokens. That keeps attention cost sub-quadratic at 200K, which is what makes the window economical to serve. The MoE routing uses 256 routed experts plus 1 shared expert, with 8+1 active per token across 78 layers.

Speculative decoding and long-horizon training

GLM-5.1 ships multi-token-prediction (MTP) speculative decoding, exposed on SGLang through EAGLE flags and on vLLM through the MTP speculative-config. Its headline capability is not a single benchmark but duration: Z.ai built an asynchronous agent RL infrastructure so the model sustains a single task for up to 8 hours, running planning, execution, testing, fixing, and delivery across hundreds of rounds and thousands of tool calls. That is the trait people keep it around for.

Benchmarks vs the Frontier

At its April 2026 launch, GLM-5.1 was the top-scoring model on SWE-bench Pro at 58.4, ahead of GPT-5.4 (57.7) and Claude Opus 4.6 (57.3). Z.ai reported it as "overall aligned with Claude Opus 4.6" across the suite. The scores below are Z.ai's self-reported figures from the GLM-5.1 technical report.

Two calibration notes. First, Z.ai reports SWE-bench Pro, not SWE-bench Verified, so do not compare this 58.4 to a Verified number from another model. Second, Terminal-Bench here is version 2.0 (63.5), which rises to 66.5 with Claude Code as the scaffold; harness choice moves that figure by several points. See our SWE-bench Pro breakdown for how these harnesses behave and where scaffolds inflate scores.

Serving Footguns: SGLang and vLLM

The most common "GLM-5.1 is broken" report is a parser mismatch, not a model problem. GLM-5.1 uses the glm47 tool-call format. If your server is configured with the older glm45 tool-call parser, the call is left as raw text in message.content and never surfaces as a function call. On both SGLang and vLLM you need --tool-call-parser glm47 with the reasoning parser set to --reasoning-parser glm45.

# SGLang, single 8x H200 node, FP8
python -m sglang.launch_server \
  --model-path zai-org/GLM-5.1 \
  --tp 8 \
  --tool-call-parser glm47 \
  --reasoning-parser glm45 \
  --speculative-algorithm EAGLE \
  --speculative-num-steps 3 \
  --speculative-eagle-topk 1

Two more footguns worth knowing before you put it in an agent loop:

• Tool calls can drop near the context ceiling. A known vLLM issue (#42400) has GLM-5.1-FP8 tool-call parsing fail intermittently as the context approaches the 202,752 max_model_len, typically when Claude Code enters planning mode and emits a complex structured tool call while MTP speculative decoding is on. If you see intermittent tool failures only on long sessions, this is why.
• MTP and tool calling together want a recent build. Running MTP speculative decoding and tool calling at the same time works best on current SGLang (v0.5.10+) and vLLM (v0.19.0+); older builds handle one path cleanly but can trip on the combination.

Self-host memory reality

At 754B total the BF16 footprint is about 1.5TB, so single-node inference means the FP8 build (about 754GB), which fits on 8x H200 SXM5 with headroom at --tp 8. An INT4 (AWQ) quant drops to about 377GB plus 5-10% for KV cache and activation buffers. For a rough per-node math: FP8 weights plus KV cache for a full 200K context is the constraint, not raw parameter count, since only 40B are active per token.

GLM-5.1 API: Pricing by Provider

GLM-5.1 is available through an OpenAI-compatible API on multiple hosts. Z.ai's first-party list pricing is $1.40/M input, $0.26/M cached input, and $4.40/M output, the same sticker as GLM-5.2. Third-party providers serve the same weights cheaper: DeepInfra lists $1.05/$3.50 and OpenRouter routes near $0.97/$3.04. The variation reflects how each host serves activations (FP8 quantization moves output away from the reference weights) and how they price margin, not a difference in the model.

Sources: Z.ai list pricing from the Z.ai pricing docs; DeepInfra from its GLM-5.1 pricing guide; OpenRouter from openrouter.ai/z-ai/glm-5.1; Together from its GLM-5.1 model page. Marketplace routers vary by upstream host and quantization, so the cheaper sticker can mean FP8-quantized activations rather than reference weights.

from openai import OpenAI

client = OpenAI(
    base_url="https://api.z.ai/api/paas/v4",
    api_key="YOUR_ZAI_KEY",  # Authorization: Bearer <key>
)

resp = client.chat.completions.create(
    model="glm-5.1",
    messages=[
        {"role": "user", "content": "Explain this stack trace and propose a fix."},
    ],
)
print(resp.choices[0].message.content)

To run the current-generation model on managed infrastructure, Morph serves morph-glm52-744b (the GLM-5.2 successor) at $1.10/M input and $4.10/M output with a 1M context, on custom kernels via the same OpenAI-compatible endpoint as every other Morph model:

from openai import OpenAI

client = OpenAI(
    base_url="https://api.morphllm.com/v1",
    api_key="YOUR_MORPH_API_KEY",
)

resp = client.chat.completions.create(
    model="morph-glm52-744b",
    messages=[
        {"role": "user", "content": "Refactor this function to remove the nested loop."},
    ],
)
print(resp.choices[0].message.content)

Use GLM-5.1 in Claude Code and Cline

Claude Code

Z.ai's coding API speaks the Anthropic format, so Claude Code needs only environment variables. Point the Sonnet/Opus aliases at glm-5.1:

export ANTHROPIC_BASE_URL="https://api.z.ai/api/anthropic"
export ANTHROPIC_AUTH_TOKEN="your-z-ai-key"
export ANTHROPIC_DEFAULT_OPUS_MODEL="glm-5.1"
export ANTHROPIC_DEFAULT_SONNET_MODEL="glm-5.1"
claude

For routing GLM-5.1 alongside other providers behind one endpoint, see Claude Code with LiteLLM.

Cline

Cline uses an OpenAI-compatible provider. Set the base URL to https://api.z.ai/api/coding/paas/v4, the model to glm-5.1, and the context window to 200000. Leave image support unchecked; GLM-5.1 is text-only. To point Cline at Morph's current model instead, use base URL https://api.morphllm.com/v1 and model morph-glm52-744b.

GLM-5.1 vs GLM-5.2: Should You Upgrade?

GLM-5.2 is the better model on paper: 62.1 vs 58.4 on SWE-bench Pro, a 1M-token context versus 200K, IndexShare attention, and dual thinking modes. But the upgrade is not free, and for a real slice of GLM-5.1 users it is not worth it yet. Here is the honest split.

One caveat on GLM-5.2 to weigh before moving: Z.ai disclosed it exhibits more reward-hacking behavior than GLM-5.1 (reaching for reference solutions and hidden tests during RL training) and added a two-stage anti-hack module in response. If you run agents with filesystem or test access, GLM-5.1's baseline behavior is one less thing to eval-gate. Full detail on the successor is in our GLM-5.2 breakdown.

GLM-5.1: Pros and Cons

When to Use GLM-5.1 (and When Not)

Frequently Asked Questions

What is GLM-5.1?

An open-weight (MIT) mixture-of-experts coding model from Z.ai (formerly Zhipu AI), released April 7, 2026. 754B total parameters (40B active), 200K-token context, 128K max output. It led open-weights on SWE-bench Pro at launch (58.4). Weights on Hugging Face (zai-org/GLM-5.1). Its successor GLM-5.2 shipped June 2026.

What is GLM-5.1's context window?

200K tokens (202,752 max_model_len) with up to 128K-131K max output. That covers most single-repository codebases in one pass. GLM-5.2 raised this to 1M via IndexShare, so if you routinely exceed 200K, that is the reason to upgrade.

How much does GLM-5.1 cost?

Z.ai list pricing: $1.40/M input, $0.26/M cached, $4.40/M output. Third-party hosts are cheaper: DeepInfra $1.05/$3.50, OpenRouter routed near $0.97/$3.04. Because GLM-5.1 is less verbose than GLM-5.2, per-task cost is often lower even at the same sticker.

GLM-5.1 vs GLM-5.2: which should I use?

GLM-5.2 wins on benchmarks (62.1 vs 58.4 SWE-bench Pro) and context (1M vs 200K). GLM-5.1 wins on stability and cost: a serving stack settled since April, cheaper hosting, and less runaway verbosity. Upgrade if you hit the 200K limit, need output past 32K, or want the top score. Stay on 5.1 if your work fits in 200K and you want fast, cheap iteration.

Why does my GLM-5.1 tool calling fail?

GLM-5.1 uses the glm47 tool-call format. If your server uses the older glm45 parser, the call stays as raw text in message.content and is never parsed. Use --tool-call-parser glm47 with --reasoning-parser glm45. There is also a known vLLM issue (#42400) where parsing drops near the 200K ceiling with MTP on.

Can I self-host GLM-5.1?

Yes, under MIT. At 754B total the BF16 footprint is ~1.5TB. The FP8 build is ~754GB and fits on 8x H200 SXM5 with headroom at --tp 8; an INT4 (AWQ) quant is ~377GB. Weights are at zai-org/GLM-5.1 on Hugging Face.

Is GLM-5.1 still worth using now that GLM-5.2 is out?

Yes, for a specific profile: work that fits in 200K context, tight iteration loops where 5.1's lower verbosity is faster and cheaper, and setups where a settled serving stack and cheaper third-party hosting matter more than the last few benchmark points. If you need the 1M window or the top score, move to GLM-5.2.