By ByteBot
DeepSeek released V4.1 Flash on September 10 under an MIT license — a 552-billion-parameter mixture-of-experts model that costs $0.15 per million tokens off-peak and outscores Claude Opus 5 and GPT-5.6 Sol on several coding benchmarks. The weights are on Hugging Face. Frontier-adjacent performance, commercially usable open weights, and pricing that makes Opus 5 look like it charges for the building lease. That is the story.
What Actually Changed
V4.1 Flash is not a typical “Flash” tier model. DeepSeek’s new Causal Encoder-Decoder (CED) architecture runs 40 transformer layers split into a 20-layer encoder followed by a 20-layer decoder. Total parameter count is 552B, but the mixture-of-experts routing activates roughly 8B parameters during prefill and 16B during decode — so you get heavyweight capacity at lightweight compute cost per token.
The KV cache is now one-eighth the SSD footprint of the prior generation, which matters a lot if you are running agentic pipelines that re-read long contexts repeatedly. Native multimodal vision is baked in: no separate endpoint, no vision-specific model name. Images and text go in together. The 1-million-token context window stays.
MIT License: The Part That Actually Matters
Previous DeepSeek open-weight releases came with commercial restrictions. V4.1 Flash does not. MIT means commercial use, redistribution, and fine-tuning with no revenue thresholds and no enterprise agreements required. If your workload involves sensitive data you cannot send to a third-party API, you can now run a frontier-class model entirely on your own hardware.
The hardware floor is real: the checkpoint weighs 510 GB across 48 files on Hugging Face. Minimum self-hosting requires 8x H100 GPUs (FP8 build). FP4 quantization drops the footprint to roughly 385 GiB, and a single B200 covers the native FP4+FP8 build. SGLang has a V4.1 preview with an opt-in that keeps the Engram memory tables in host RAM rather than GPU memory. Full vLLM support is still in progress.
Pricing: Schedule Around Peak Hours
The model name is deepseek-flash. The pricing structure rewards workloads that can run off-peak.
| Tier | Cache Hit Input | Cache Miss Input | Output |
|---|---|---|---|
| Off-peak | $0.003/M | $0.15/M | $0.60/M |
| Peak | $0.006/M | $0.30/M | $1.20/M |
Peak hours are 01:00–04:00 and 06:00–10:00 UTC, Monday through Friday. At off-peak rates, output costs roughly 40 times less than Claude Opus 5. Cache hits off-peak hit $0.003 per million — 98% cheaper than a cache miss. If you are running long-context agentic tasks with repeated context re-reads, that cache-hit rate is the number that changes your infrastructure math.
Benchmarks: Impressive With Caveats
DeepSeek’s published numbers are striking. CyberGym: 88.1 versus GPT-5.6 Sol’s 84.5. Codeforces rating: 3471 versus the previous V4 Pro’s 3348. Terminal-Bench 3.0: 30.0 versus V4-Pro’s 11.8. Independent tests do not fully reproduce the launch benchmarks — that is worth stating clearly. But even partial replication at 40x lower cost than Opus 5 changes the evaluation calculus. This is worth testing even if you trust Anthropic or OpenAI for production.
The Migration Problem
DeepSeek retired deepseek-v4-flash and deepseek-v4-flash-vision-exp with roughly one day’s notice. Beginning September 14 at 04:00 UTC, deepseek-v4-pro also routes to V4.1 Flash — billed at Flash rates. There was no parallel migration window. Check the official DeepSeek API changelog for the latest routing details.
The code change is trivial:
# Before (deprecated, temporarily aliased)
model = "deepseek-v4-flash"
# After
model = "deepseek-flash"The disruption is not the rename — it is that teams running automated evals or prompt regression tests against the old endpoint names now have untracked model drift. Output behavior, latency, and formatting can all shift when a different underlying model starts answering requests. Rerun your test suites.
What to Do Now
- Update model names:
deepseek-v4-flash→deepseek-flash - Rerun eval and regression suites — assume nothing carries over from
deepseek-v4-probehavior - Test off-peak scheduling for batch or agentic workloads; the cache-hit pricing alone may justify restructuring job queues
The MIT license and the pricing are real. Whether the benchmarks hold up under independent testing is still an open question. But the open-weights story here is not theoretical — this is one of the first times a model at this capability level has come with no commercial strings attached. That changes the calculus for privacy-sensitive and regulated workloads that previously had no viable open-weight option. Here is a self-hosting hardware guide if you want to evaluate local deployment.
