Baseten Claims Fastest GLM-5 Speeds on Artificial Analysis Benchmark

What Baseten reports

Baseten, an inference provider, says Artificial Analysis independently benchmarked its Model API for GLM-5 and found state-of-the-art results for both time to first token (TTFT) and tokens per second (TPS). The company's screenshot of the result shows 200+ TPS and about 200ms TTFT, which Baseten describes as the fastest on the leaderboard.

GLM-5 is the latest open-weight flagship from Z.ai. According to Baseten, it has 744B total parameters, 40B active per forward pass, was trained on 28.5T tokens and is released under the MIT license. Baseten calls it the most capable open-source model in the world on coding and agentic benchmarks, citing 77.8 on SWE-bench Verified, 92.7% on AIME 2026 and first place among open models on Vending Bench 2.

Because GLM-5 is a reasoning model, it produces a thinking sequence before its final answer. Baseten argues this makes tokens per second the most important number for production use, since the thinking sequence is often longer than the final response.

The three optimizations behind the speed

Baseten says it combined its Inference Stack, which includes KV-aware routing, with three model-specific changes.

• MoE kernels: GLM-5 shares its Mixture of Experts lineage with DeepSeek V3 but uses more layers and a smaller hidden size. Baseten says standard MoE kernels profiled against DeepSeek V3's dimensions do not run in their best regime on GLM-5, so it tuned its dispatch kernels to the actual shape of the model.
• DeepSeek Sparse Attention (DSA): GLM-5 is the first GLM model to use DSA, which Baseten says sparsifies attention across the sequence and reduces KV cache footprint at long context while keeping the full 200K context window. A naive implementation falls back to dense attention and loses most of the memory benefit, so Baseten built kernels that exploit the sparse pattern directly. The company says this allows larger batch sizes and better throughput under load.
• MTP speculative decoding: GLM-5 ships with native Multi-Token Prediction heads, so no separate draft model has to be trained or hosted. Baseten says its speculative decoding engine keeps host-side coordination lightweight and integrates draft scheduling into the inference loop, achieving high token acceptance rates on code generation, long-horizon reasoning and agentic tool use.

Why speed matters for reasoning models

Speculative decoding works by generating candidate tokens cheaply and verifying them in parallel with the target model. When the target accepts the draft tokens, several tokens are produced for the cost of one forward pass. Baseten says the approach improves tokens per second without sacrificing quality, but only if the overhead of managing draft tokens stays small.

For a reasoning model that can generate tens of thousands of thinking tokens before answering, Baseten says higher TPS translates directly and linearly into faster end-to-end response time.

What it means for developers

Baseten positions the result for teams building autonomous coding agents, multi-step tool-use pipelines and other applications that need frontier-grade intelligence at production speed. The company says that with 77.8 on SWE-bench Verified, GLM-5 approaches Claude Opus 4.5 on software engineering, and that at 200+ tokens per second on Baseten, GLM-5-based applications run significantly faster than Claude while remaining fully open-weight under MIT.

These performance and comparison claims come from Baseten. The Artificial Analysis benchmark is described as independent, but the source does not provide the underlying benchmark data beyond the company's screenshot summary, and it does not give pricing or availability details.