Uploaded August 2026 | Updated September 2026, 2 weeks ago
DeepSeek V4 Flash vs Qwen 3.8 Flash Next: 1-Bit Local Benchmark
DeepSeek V4 Flash is a 284B model, and its 1-bit build is 76.87 GB, which fits inside a desktop with 96 GB of RAM. In this video it runs against Qwen 3.8 Flash Next on 19 hard problems, first on the CPU alone and then with expert layers offloaded to an RTX 5090.
We start with the architecture, because the speed numbers only make sense once you can see where the parameters live: 256 wide experts across 43 DeepSeek layers, 512 tiny experts across 48 Qwen blocks, a 51B n-gram table, and 13B versus 6B active per token. Then we measure both models on the same machine and read what each one actually produced. Everything you need is in this video, so you can follow along on your own hardware from start to finish.
📺 Full playlist: youtube.com/playlist?list=PLc2rvfiptPSReropGbvDFpB6dneNBwqhD
⏱ Chapters:
0:00 Intro and what we are testing
1:09 DeepSeek V4 Pro vs Flash: what fits in 96 GB
3:13 Qwen 3.8 Flash Next: 180B with 6B active
4:54 Spec table: 284B and 180B side by side
5:52 What a decoder block actually holds
8:49 Expert inner width and the FFN split into experts
13:00 How many experts fire for one token
15:53 Parameters inside a single DeepSeek expert
19:19 Adding it up to 277B across 43 layers
20:39 Where the last 7B comes from
22:44 Wide experts vs tiny experts
23:35 Counting Qwen's 180B: n-gram table and draft head
26:27 Active parameters: 13B vs 6B per token
27:57 Test setup: Core Ultra 9 285K, RTX 5090, 96 GB RAM
28:30 Results on 19 problems: 16 vs 13
29:25 How each model fails
31:01 The three.js game task
32:26 What 1-bit quantization costs
33:56 Speed: CPU only vs GPU offload
🔗 Resources:
Full write-up with every table and the raw completions: kgptalkie.com/tutorials/llm-benchmarking/deepseek-v4-flash-1bit-vs-qwen-flash-next
Both 1-bit builds are Unsloth UD-IQ1_S: huggingface.co/unsloth
Rig: Intel Core Ultra 9 285K, RTX 5090 with 32,607 MB VRAM, 95.3 GB system RAM, Windows 11 Pro
The two llama-server flags that mattered most: --n-cpu-moe 30 (keeps 30 of DeepSeek's 43 expert layers on the CPU) and --parallel 1 (frees the KV cache that three unused slots were holding, worth 14.6 to 23.5 tokens a second on its own)
📺 Watch next:
I Ran Qwen 3.8 Flash Next (Qwen 4) on CPU Only: youtube.com/watch?v=G73xv1ARjeI
Qwen 3.8 Flash Next (Qwen 4) vs 27B Architecture Teardown: youtube.com/watch?v=VvV3QRHfzYU
🎓 Go deeper, my Udemy course:
Fine Tuning LLM with Hugging Face Transformers for NLP: kgptalkie.com/fine-tuning-llm
Give this a like if the parameter math finally clicked, and tell me in the comments which model you want on the bench next and how much RAM you are working with. I read every one. Subscribe and hit the bell to catch the rest of the local LLM series.
#DeepSeek #Qwen #LocalLLM #LLM #AI
DeepSeek V4 Flash vs Qwen 3.8 Flash Next: 1-Bit Local Benchmark
DeepSeek V4 Flash is a 284B model, and its 1-bit build is 76.87 GB, which fits inside a desktop with 96 GB of RAM. In this video it runs against Qwen 3.8 Flash Next on 19 hard problems, first on the CPU alone and then with expert layers offloaded to an RTX 5090.
We start with the architecture, because the speed numbers only make sense once you can see where the parameters live: 256 wide experts across 43 DeepSeek layers, 512 tiny experts across 48 Qwen blocks, a 51B n-gram table, and 13B versus 6B active per token. Then we measure both models on the same machine and read what each one actually produced. Everything you need is in this video, so you can follow along on your own hardware from start to finish.
📺 Full playlist: youtube.com/playlist?list=PLc2rvfiptPSReropGbvDFpB6dneNBwqhD
⏱ Chapters:
0:00 Intro and what we are testing
1:09 DeepSeek V4 Pro vs Flash: what fits in 96 GB
3:13 Qwen 3.8 Flash Next: 180B with 6B active
4:54 Spec table: 284B and 180B side by side
5:52 What a decoder block actually holds
8:49 Expert inner width and the FFN split into experts
13:00 How many experts fire for one token
15:53 Parameters inside a single DeepSeek expert
19:19 Adding it up to 277B across 43 layers
20:39 Where the last 7B comes from
22:44 Wide experts vs tiny experts
23:35 Counting Qwen's 180B: n-gram table and draft head
26:27 Active parameters: 13B vs 6B per token
27:57 Test setup: Core Ultra 9 285K, RTX 5090, 96 GB RAM
28:30 Results on 19 problems: 16 vs 13
29:25 How each model fails
31:01 The three.js game task
32:26 What 1-bit quantization costs
33:56 Speed: CPU only vs GPU offload
🔗 Resources:
Full write-up with every table and the raw completions: kgptalkie.com/tutorials/llm-benchmarking/deepseek-v4-flash-1bit-vs-qwen-flash-next
Both 1-bit builds are Unsloth UD-IQ1_S: huggingface.co/unsloth
Rig: Intel Core Ultra 9 285K, RTX 5090 with 32,607 MB VRAM, 95.3 GB system RAM, Windows 11 Pro
The two llama-server flags that mattered most: --n-cpu-moe 30 (keeps 30 of DeepSeek's 43 expert layers on the CPU) and --parallel 1 (frees the KV cache that three unused slots were holding, worth 14.6 to 23.5 tokens a second on its own)
📺 Watch next:
I Ran Qwen 3.8 Flash Next (Qwen 4) on CPU Only: youtube.com/watch?v=G73xv1ARjeI
Qwen 3.8 Flash Next (Qwen 4) vs 27B Architecture Teardown: youtube.com/watch?v=VvV3QRHfzYU
🎓 Go deeper, my Udemy course:
Fine Tuning LLM with Hugging Face Transformers for NLP: kgptalkie.com/fine-tuning-llm
Give this a like if the parameter math finally clicked, and tell me in the comments which model you want on the bench next and how much RAM you are working with. I read every one. Subscribe and hit the bell to catch the rest of the local LLM series.
#DeepSeek #Qwen #LocalLLM #LLM #AI
