Question 1

What is new in the December 2025 Open-Prem update?

Accepted Answer

The December 2025 update covers two landmark releases that achieved frontier parity with proprietary systems: DeepSeek V3.2 and Zhipu AI's GLM-4.6. For the first time, open-source models matched the performance of the most advanced proprietary systems including GPT-5 and Gemini-3.0-Pro, and in some cases surpassed them. Both models are MIT-licensed, enabling complete enterprise control.

Question 2

What is DeepSeek V3.2 and why does it matter?

Accepted Answer

DeepSeek V3.2, released December 1, 2025, rivals GPT-5 and Google Gemini-3.0-Pro. It introduces DeepSeek Sparse Attention reducing computational complexity by 70%. The model has 685 billion total parameters with 37 billion active per token, a 128K token context window, and MIT licensing. DeepSeek V3.2-Speciale achieved gold medals at the 2025 International Mathematical Olympiad (35/42 points) and International Olympiad in Informatics (492/600 points).

Question 3

What is GLM-4.6?

Accepted Answer

GLM-4.6, released September 30, 2025 by Zhipu AI, is the only frontier-scale model (355 billion parameters) available under MIT license. It achieves a 48.6% win rate against Claude Sonnet 4 on human-evaluated coding tasks, has a 200K token context window, and supports native tool-calling and hybrid reasoning modes. It enables enterprises to self-host, customize, and own their AI infrastructure without vendor lock-in.

Question 4

What does frontier parity mean for enterprise AI strategy?

Accepted Answer

Frontier parity means open-source models now match the most advanced proprietary systems in performance. The December 2025 update shows DeepSeek V3.2 matching GPT-5 on AIME 2025 math benchmarks (93.1% vs 94.6%) and surpassing it on SWE Multilingual (70.2% vs 55.3%). This means organizations no longer sacrifice capability by choosing open-source, and can achieve 80-90% cost savings at API level or 95%+ savings with self-hosting at scale.

Question 5

How much does open-prem deployment cost compared to proprietary APIs in December 2025?

Accepted Answer

The December 2025 update provides direct cost comparisons: Claude Sonnet 4.5 costs $3.00/$15.00 per million input/output tokens. GPT-5 costs $2.50/$10.00. GLM-4.6 via API costs $0.60/$2.00. DeepSeek V3.2 via API costs $0.27/$1.10. Self-hosted GLM-4.6 costs $0.05-$0.20 per million tokens. This represents savings of 80-90% at API level and 95%+ with self-hosting.

Question 6

What hardware is needed to self-host GLM-4.6?

Accepted Answer

GLM-4.6 requires 8 H100 GPUs or 4 H200 GPUs for standard inference at FP8 precision, 16 H100 or 8 H200 GPUs for full 128K context, 1TB+ system RAM, and 400GB disk space for the full model or 135GB for quantized GGUF format. Community quantizations including Unsloth Dynamic 2-bit GGUF achieve 75% size reduction with minimal accuracy loss.

Question 7

What is DeepSeek Sparse Attention?

Accepted Answer

DeepSeek Sparse Attention (DSA) is a breakthrough in the V3.2 architecture that reduces computational complexity from O(L-squared) to O(kL), where L is sequence length and k is selected tokens. Instead of processing all tokens equally, DSA uses a lightning indexer and fine-grained token selection to identify and process only the most relevant information, achieving a 70% reduction in inference costs compared to the previous version.

Question 8

What is the role of Chinese AI labs in the Open-Prem ecosystem?

Accepted Answer

The December 2025 update reports that Chinese open-source AI models now account for 17% of global downloads. DeepSeek and Zhipu AI (GLM) are both Chinese labs releasing frontier-quality models under MIT license. The paper notes this reflects a strategic response to US export controls that is reshaping the global AI landscape, and recommends enterprises consider vendor diversification to reduce dependence on any single AI provider or geographic region.

DAVID BORISH

THE OPEN-PREM
INFLECTION POINT V2
(DECEMBER 2025 UPDATE)

When Open-Source LLMs Achieve Frontier Parity

WHAT CHANGED IN V2: DECEMBER 2025 UPDATE

ABOUT THE AUTHOR

FREQUENTLY ASKED QUESTIONS

What is new in the December 2025 Open-Prem update?

What is DeepSeek V3.2 and why does it matter?

What is GLM-4.6?

What does frontier parity mean for enterprise AI strategy?

How much does open-prem deployment cost compared to proprietary APIs in December 2025?

What hardware is needed to self-host GLM-4.6?

What is DeepSeek Sparse Attention?

What is the role of Chinese AI labs in the Open-Prem ecosystem?

Back to top

THE OPEN-PREM INFLECTION POINT V2 (DECEMBER 2025 UPDATE)

When Open-Source LLMs Achieve Frontier Parity

WHAT CHANGED IN V2: DECEMBER 2025 UPDATE

ABOUT THE AUTHOR

​

FREQUENTLY ASKED QUESTIONS

What is new in the December 2025 Open-Prem update?

What is DeepSeek V3.2 and why does it matter?

What is GLM-4.6?

What does frontier parity mean for enterprise AI strategy?

How much does open-prem deployment cost compared to proprietary APIs in December 2025?

What hardware is needed to self-host GLM-4.6?

What is DeepSeek Sparse Attention?

What is the role of Chinese AI labs in the Open-Prem ecosystem?

Back to top

THE OPEN-PREM
INFLECTION POINT V2
(DECEMBER 2025 UPDATE)