Diffusion models usually trade familiar left-to-right generation for the ability to refine multiple tokens together. The important question is whether that parallelism survives at a scale useful for real agent work. Reported results suggest it can: the model pairs blockwise denoising with a sparse MoE backbone, though its speed claims depend on specific GPUs and workloads.
Key Capabilities
- Blockwise generation: Tokens within a block can attend bidirectionally and be resolved in parallel, while the completed block joins a causal prefix. This changes the generation process rather than merely accelerating standard autoregressive decoding.
- Sparse scale: The Qwen3.6-35B-A3B-derived backbone has 35B total parameters and roughly 3B active per token. That keeps per-token computation lower than a dense model of similar total size, while retaining a large expert pool.
- Long context and agent results: The model supports up to 131,072 tokens. DiffuSpace reports 73.6 on SWE-bench Verified and 51.7 on Terminal-Bench 2.1; these are vendor-reported benchmark results, not independent guarantees of performance on every coding workflow.
- Hardware-dependent throughput: DiffuSpace reports 680 tokens/s on a GB300 and 515 tokens/s on an H200 for agent workloads. These figures are tied to the stated hardware and internal evaluation setup, so they should not be treated as universal latency expectations.
Who It’s For
A good fit for researchers and developers testing diffusion-based generation, especially on coding or tool-use tasks, who can provide substantial GPU memory and work with evolving inference support. Look elsewhere if you need a lightweight local model, predictable autoregressive decoding, or performance evidence from independent evaluations; BF16 weights alone require about 69 GB, before inference state.
