DeepSeek’s Post-Training Lead Declares “Entering the Era of Democratized Wish-Driven Coding” As V4.1-Flash-1017 Model Smokes OpenAI’s GPT-5.6 Sol

"DeepSeek-V4.1-Flash-1017 scores 77 percent on DeepSWE compared to 74 percent by Opus 5.0, 60.5 percent on AutomationBench against 50.3 percent, and 72.1 percent on Apex Bench against 65.8 percent.

DeepSeek has retained the core architecture of the previous version of the V4.1-Flash - including Causal Encoder-Decoder (CED), Compressed Sparse Attention 2 (CSA2), FP4 Quantization, and SWA Bounded Replay - that had played an important role in reducing the variant's overall KV cache to just 890 bytes per token!

DeepSeek-V4.1-Flash-1017, however, also introduces the following tweaks:

1. Harness Optimization: It has been hardwired to work seamlessly with agentic coding environments, including DeepSeek's own harness. If you put this AI into a loop, where it tests its own code, fails, fixes it, and tries again, it performs significantly better than older version.
2. Extreme Output Condensation: DeepSeek specifically trained this model to stop being "wordy." Instead of vomiting paragraphs of explanations, it is tuned to give you the exact code or answer you need immediately. Shorter answers mean fewer tokens, which directly equals faster speeds and much cheaper costs.

DeepSeek also paid particular attention to reinforcement learning (RL), which is akin to training a dog by using rewards and punishments. In this case, the RL system penalized the model for rambling, and rewarded it for finding the shortest, most efficient path to the correct answer.


All of these steps mean that DeepSeek-V4.1-Flash-1017 now delivers the maximum bang for the proverbial buck, delivering unparalleled performance at minimal cost."

Source: https://wccftech.com/deepseeks-post...-flash-1017-model-smokes-openais-gpt-5-5-sol/

And best of all you can run it locally at good speed (with the right setup).

763Billion parameter at good speed that will be quite costly and risky investment (you built all of this and 3 months later something so much better from anthropic-openai come up running on Rubin/Groq your competition use..), ability to run them on the cloud but your own locked private AWS/google/other servers is also nice.

LukeTbk said:

763Billion parameter at good speed that will be quite costly and risky investment (you built all of this and 3 months later something so much better from anthropic-openai come up running on Rubin/Groq your competition use..), ability to run them on the cloud but your own locked private AWS/google/other servers is also nice.

Click to expand...

Already got 4 DGX Sparks on my desk so it's just a case of loading on the latest and greatest LLMs as they are released. It is quite nice to have that ability.

mashie said:

Already got 4 DGX Sparks on my desk so it's just a case of loading on the latest and greatest LLMs as they are released. It is quite nice to have that ability.

Click to expand...

#Jealous