DeepSeek V4 Pro 0813 1.6T vs Qwen3.8 2.4T-A95B

Analysis

DeepSeek V4 Pro 0813 and Qwen3.8 2.4T-A95B are both frontier-scale, text-only, datacenter-only MoE releases from August 2026, and Qwen is unambiguously the bigger model on paper: 53% more total parameters, nearly double the active count. On the one benchmark both report on a matching suite label, Terminal-Bench 2.1, the smaller model wins. And on realistic single-machine deployment, the smaller model loses just as clearly, for a reason that has nothing to do with its size.

DeepSeek V4 Pro 0813 is 1.6T total parameters with 49B active, spread evenly across 61 layers that each use the same hybrid Compressed Sparse Attention/Heavily Compressed Attention design, compressing every layer's KV cache roughly 10x. Qwen3.8 2.4T-A95B is 2,446.18B total with 95B active, and hybrid in a different way: of its 92 layers, only 23 keep a real KV cache (4 KV heads x 256 head dim), the other 69 running Gated DeltaNet, a linear-attention design with a fixed-size state instead of a cache. Qwen's native context window ends at 262,144 tokens, a quarter of DeepSeek's 1,048,576; restricted to that shared ceiling, DeepSeek's KV cache is 3.3 GB against Qwen's 24.7 GB, DeepSeek's compression winning out over Qwen's smaller caching-layer count the same way it does against Kimi K3. Weights track total parameter count almost exactly: at each model's own recommended quantization, DeepSeek needs about 609.6 GB, Qwen about 932.0 GB, and the full picture at the shared 262,144-token context is 686.4 GB for DeepSeek against 1,071.5 GB for Qwen, in line with Qwen's 53% larger parameter count. On Terminal-Bench 2.1, the benchmark both models report, DeepSeek actually leads despite being the smaller model: 87.9 against Qwen's 86.6, DeepSeek's own comparison run against Alibaba's own reported figure. Qwen publishes a broader academic suite DeepSeek hasn't republished for the 0813 checkpoint: GPQA Diamond 92.6 and SWE-bench Pro 67.7, with nothing directly comparable on DeepSeek's side since it stopped reporting those fields after the April preview. Licensing favors DeepSeek: unrestricted MIT, against Qwen's custom Qwen3.8-Max license, free for commercial use under $20M/month revenue or 100M monthly active users (with attribution required above that), and a separate license required for a Model-as-a-Service or AI-assistant business over $50M/year. Neither model accepts image or video input; both are text-only. The more surprising asymmetry shows up in how each model is actually reachable at all: Qwen3.8 has a real, if extreme, single-machine path, Unsloth's UD-Q1_0 GGUF at roughly 397 GB, small enough to fit inside a 512 GB Mac Studio M3 Ultra's unified memory with room for context and overhead. DeepSeek V4 Pro 0813 has no equivalent: no independently vetted low-bit build exists yet, only unverified community conversions, so the generic Q2_K ladder estimate, 609.6 GB of weights alone, is the only number available, and it doesn't fit that same 512 GB machine. The larger model, in other words, is currently the more realistically self-hostable one.

Bottom line: Despite being the smaller model on paper, DeepSeek V4 Pro 0813 is not the more practical local option here; Qwen3.8 2.4T-A95B is, purely because Qwen has a real sub-2-bit GGUF build and DeepSeek doesn't yet. Anyone with a 512 GB Mac Studio M3 Ultra (or a comparable multi-GPU pool) who specifically needs open weights, for auditability, fine-tuning, or an air-gapped deployment, should look at Qwen3.8's UD-Q1_0 build before DeepSeek's. For everyone else living in the API, the two are close: DeepSeek slightly leads the one shared benchmark, carries the cleaner license, and needs meaningfully less context-scaling memory if a workload actually reaches into the hundreds of thousands of tokens, while Qwen brings a fuller published academic benchmark record and a 53%-larger active-parameter budget that may matter for harder reasoning tasks the shared benchmark doesn't capture.