GLM-5.3 Flash vs DeepSeek-V4.1 Flash: 20% Faster Code Gen on M5 Ultra

Benchmark Source and Setup

Overseas engineer Nilesh (X user @nileshgr) published a benchmark on October 30, 2026, running two open-source MoE models on the same Mac Studio M5 Ultra 256GB : GLM-5.3 Flash and DeepSeek-V4.1-Flash 1017 . Test conditions: identical machine, identical prompts, fixed random seed (42), thinking disabled, prompt cache disabled, response truncated to 512 tokens, and 8 concurrent requests.

The raw screenshot from the original tweet is shown below:

The author extracted the numbers into a comparison table:

1. Title Claims "2x Faster" but Data Shows ~1.2x

The tweet title stated "GLM-5.3 Flash decode code 2x faster" and Chinese text echoed "快 2 倍". Actual measurements:

Code decoding: GLM-5.3 Flash 84.9 tok/s vs DeepSeek 71.2 tok/s = 1.19x faster.

Plain text: 71.2 vs 68.7 tok/s = only 3.7% faster.

The author's own chart also labels the gain as 1.2x .

Conclusion: the "2x" claim is clickbait; the real gap is approximately 20% . Still, a 20% gain on a 256GB unified-memory MoE is meaningful.

2. Real Gap Widens Under Concurrency: 8 Parallel Requests

Single-stream decoding differs by only 20%, but with 8 simultaneous requests GLM-5.3 Flash delivers 109.2 tok/s total throughput versus DeepSeek's 93.6 tok/s . For workflows that run multiple agents in parallel (e.g., several coding agents working concurrently), this difference compounds.

3. First-Token Latency on Long Prompts: Mixed Results

The tweet claimed "DeepSeek gets to the first token sooner on long prompts", but the data reveals a crossover:

16K prompt: DeepSeek 7.7s vs GLM 8.2s → DeepSeek faster.

32K prompt: 16.1s vs 16.9s → DeepSeek slightly faster.

64K prompt: 31.5s vs 30.9s → essentially tied, GLM edges ahead.

128K prompt: 68.0s vs 61.2s → GLM faster by ~10%.

Thus, DeepSeek leads at 16K-32K context, while GLM wins at 64K-128K . The optimal choice depends on your actual prompt lengths.

4. Critical Test Condition Caveats

The comparison is not a pure model-vs-model test; several configuration differences exist:

GLM-5.3 Flash ran on TensorFold 0.5.5 with MTP (multi-token prediction) enabled (Bob's Mac recipe), using greedy decoding.

DeepSeek ran on 1Q-MLX 3.0 + mix-serve , with 0 DSpark draft, model-card default temperature 1.0.

DeepSeek's server reported numerous 1-second outliers.

Only speed was measured; output quality was not evaluated.

Different engines, speculative strategies, and decoding parameters mean the 20% gap is a mixed engine+model result , not a pure capability difference.

5. Model Backgrounds

Both are MoE architectures with huge total parameters but small active parameter counts, enabling them to fit and run at tens of tok/s in 256GB unified memory. DeepSeek-V4.1-Flash is MIT-licensed (commercial-friendly); GLM-5.3 Flash weights are available via Zhipu's open channels.

6. Decision Guide: Which Model for Your Mac?

Primary coding, single-stream speed priority → GLM-5.3 Flash slightly ahead (84.9 vs 71.2 tok/s), but note the advantage comes from TensorFold + MTP; switching engines could close the gap.

Long-document / long-context workloads → At 128K GLM is 10% faster on first token, but at 16K-32K DeepSeek is quicker. Your typical prompt length decides.

Answer quality matters → Not tested in this benchmark. The author explicitly states "speed is only one metric; quality evaluation later." Do not treat a single speed run as definitive.

Recommendation: both are top-tier open-weight MoE models that fit in 256GB. Run each for half a day with your own prompts and workflow — that beats any third-party benchmark.

References

Benchmark data and screenshots from X user @nileshgr (2026-10-30), based on the mlkaisers.com open-source benchmark project. Model specifications from Zhipu and DeepSeek official releases. Speed figures are single-machine third-party measurements; results may vary across environments.