Fireworks AI announced the DeepSeek‑V4.1‑Flash model last week. In its own benchmarks, the model reaches the same coding accuracy as GPT‑6 Astra but costs only about one‑fifteenth per task. The price reported is $0.43 for a typical DeepSWE coding task. The company says the lower cost comes from a new split‑input/output architecture and an improved KV‑cache that uses less memory.
How the model saves money
DeepSeek‑V4.1‑Flash is a 552‑billion‑parameter mixture‑of‑experts (MoE) model. It activates 8 billion parameters for input and 16 billion for output, which matches the token pattern of coding agents that read a lot of code but generate relatively short patches. In the DeepSWE test the model consumed 36.9 million input tokens and 211 thousand output tokens – a 174:1 ratio. The new KV‑cache uses a quarter of the high‑bandwidth memory (HBM) and one‑eighth of the SSD storage compared with the previous DeepSeek V4, keeping cache‑hit rates at 99.6 % and cutting the portion of the bill that comes from cache hits to 60 %.
Where it shines and where it lags
On coding‑focused benchmarks (DeepSWE) the model is on par with Astra, Gemini 3.8 Flash and Claude Opus 5 in quality – all within 0.7 points of pass@1 – but far cheaper. It is 5.5 × cheaper than Gemini 3.8 Flash, 15 × cheaper than Astra and 28 × cheaper than Opus 5. For a different workload (Terminal‑Bench 2.1) the model uses four times more output tokens than Astra yet still ends up about 12 × cheaper overall. However, on a multi‑modal reasoning benchmark (Humanity’s Last Exam) the model scores 34.52 % versus Astra’s 50.40 %, indicating weaker performance on pure academic reasoning. An “oracle router” that picks the best answer from both models would beat Astra alone by 4.4 percentage points, suggesting a mixed‑model system could be beneficial.
Getting access
Fireworks offers DeepSeek‑V4.1‑Flash through a serverless API and a dedicated capacity option for production workloads. US‑hosted endpoints for regulated industries are promised soon, and the company is planning training support for early adopters.
Why it matters
For developers building AI‑assisted coding tools, the new model makes continuous, high‑iteration agent workflows affordable, turning expensive one‑off calls into inexpensive background processes. Teams that need strong reasoning across academic subjects may still need to combine models, as DeepSeek‑V4.1‑Flash lags on that type of task.
