DeepSeek-V4.1-Flash on Fireworks: 552B MoE with split activations, 15x cheaper DeepSWE

Fireworks AI released DeepSeek-V4.1-Flash on September 14, 2026, and published benchmark results showing it matches GPT-6 Astra on the DeepSWE coding benchmark at 1/15th the cost per task. The model is a 552B-parameter mixture of experts with a new encoder-decoder split activation architecture: 8B active parameters for input processing and 16B for output generation.

Split activations and KV cache cuts

The asymmetric activation budget is designed for agentic coding workloads, which are input-heavy. In Fireworks' DeepSWE run, the model consumed 36.9M input tokens per task versus 211K output tokens, a 174:1 ratio. By spending half the compute on the input side, the architecture reduces cost where most tokens are spent.

DeepSeek also improved KV cache utilization compared to the previous DeepSeek V4 generation, using one quarter of the HBM and one eighth of the SSD storage. Fireworks reports that 99.6% of input tokens in their DeepSWE run were cache hits, and cache-hit charges accounted for 59.9% of the total cost per task. The 4x HBM reduction and 8x SSD reduction keep cache hit rates near 99% across long agent trajectories.

DeepSWE and Terminal-Bench results

On DeepSWE, DeepSeek-V4.1-Flash scored 74.34% pass@1 at $0.43 per task. GPT-6 Astra scored 74.12% at $6.524, Gemini 3.8 Flash scored 73.83% at $2.362, and Claude Opus 5 scored 73.65% at $11.838. Fireworks notes that run-to-run variation is 1.4 to 3.2 points, so the four models are in the same quality band. The cost difference is the differentiator: 5.5x below Gemini, 15x below Astra, and 28x below Opus.

On Terminal-Bench 2.1, DeepSeek-V4.1-Flash scored 86.5% accuracy versus Astra's 87.5%. Fireworks says the model uses roughly 4x more output tokens on average than Astra, thinking longer and writing more, but the combined cost per task is 12x lower.

HLE and oracle routing

On Humanity's Last Exam (HLE), DeepSeek-V4.1-Flash scored 34.52% versus Astra's 50.40%. Fireworks states the model is better suited for agentic coding than for hard scientific reasoning. An oracle router that picks the cost-optimized correct answer from both models scored 54.80%, 4.4 points above Astra alone. Fireworks explains this is possible because the two models fail on different problems, so a routed system can exceed a standalone Astra. The oracle result is an upper bound; a real router may not capture the full difference.

Fireworks positions DeepSeek-V4.1-Flash as a cost-performance baseline for autonomous software engineering, making high-iteration agent workflows viable as continuous background infrastructure. The model is available on Fireworks serverless and dedicated API, with US-hosted endpoints coming soon for regulated industries.