DeepSeek v4 tomorrow AND better than Opus 4.6??

@andrhlt

10.3K followers · usually 5.9K views a post

Views

29.2K

Likes

1.5K

Comments

52

Shares

81

Selling · SaaS › AI Tool

DeepSeekseen in 18 videos →

Why it worked

The video leverages the excitement and rapid development in the AI space by discussing leaked benchmarks and pricing for a highly anticipated model, DeepSeek v4. It taps into the community's interest in performance comparisons and cost-effectiveness.

Summary

This video discusses leaked information about DeepSeek v4, an upcoming AI model. It highlights its potential to outperform existing models like Opus 4.6 and GPT-3.5 Codex in benchmarks, while also being significantly cheaper.

Structure

1Announcement of DeepSeek v4 leak
2Comparison to Opus 4.6 and GPT-3.5 Codex
3Discussion of benchmark performance
4Analysis of pricing and parameter count
5Mention of a V4 Lite model
6Creator's personal excitement and release rumors

Product placement

DeepSeek v4 (ACTS - it is a new version of an AI model that is being discussed and benchmarked, implying its performance and features are central to the video's content. Its release and capabilities are the subject of the video. Removing it would fundamentally change the video's topic.) Opus 4.6 (ACTS - it is an AI model that is being compared to DeepSeek v4, used as a benchmark for performance. Removing it would change the comparison being made.) GPT-3.5 Codex (ACTS - it is an AI model that is being compared to DeepSeek v4, used as a benchmark for performance. Removing it would change the comparison being made.)

Original caption

DeepSeek v4 tomorrow AND better than Opus 4.6?? #claude #codex #developer #vibecode #deepseek

More from @andrhlt

Their other posts in the index, biggest breakout first.

On-screen text

DeepSeek v4 tomorrow AND better than Opus 4.6?? SWE Bench Feb 2026 (missing codex 5.3) Official SWE-bench Verified leaderboard Benchmark V4 (Leak) V3.2 V3.2-Exp Claude Opus 4.6 GPT-3.5 Codex Qwen 1.5 HumanEval (Code Gen) >90% -73.1% -67.8% 90.8% 80.0% 76.4% SWE-bench Verified 87% -73.1% -67.8% 90.8% 80.0% 76.4% Needle in a Haystack >80% 95.0 85.8 - - MMU-Pro TBD 96.0 85.8 - - GPQA Diamond TBD 82.4 91.3 - - ARC TBD 93.7 83.2 - - Codeforces Rating TBD 2188 2108 - - BrowseCoin TBD 51.6-67.6 43.1 84.0 - Evaluated with mini-gen-agent v2 mini-gen-agent.com/swebench.com V4 Flagship - Main Model Specification DeepSeek V3/V3.2 DeepSeek V4 (Leaks) Total Parameters 671B-685B Mol ~1 Trillion (1T) Mol Active Parameters/ Token ~37B ~32B (fewer despite a larger model) Context Window 128K (since Feb '26: 1M) 1 Million Tokens (native) Architecture MoE + MLA MoE + MLA + Engram Memory + n:HC + DSA Lightning Multimodal No (text only) Yes - Text, Image, Video, Audio (native) Expert Routing Top-3/Top-4 from 256 experts 16 experts active per token (from hundreds) Hardware Optimization Nvidia H800/H20 (CUDA) Huawei Ascend + Cambricon (Nvidia secondary) Training 14.8T Tokens, H800 GPUs Nvidia, inference optimized for Huawei License Input Modalities Text Text, Image, Video, Audio Output Modalities Text Text (Image/Video generation unclear) Estimated Input Price $0.28/M Tokens ~$0.14/M Tokens Estimated Output Price $0.42/M Tokens DeepSeek V4 Lite (Codename: "sealio-lite") A lighter variant has leaked alongside the flagship. At least one inference provider is testing NDA. Specification V4 Lite (Leak) Parameters ~200 Billion Context Window 1M Tokens (native) Multimodal Yes (native) Engram Memory No (according to 36kr, not integrated) vs. V3.2 "Significantly better" than current WebApp Non-Thinking vs. V3.2 Thinking Non-Thinking mode surpasses V3.2 Thinking mode Status NDA testing at inference providers We will see.

Transcript

It was leaked that DeepSeek v4 is releasing this week. Some sources say tomorrow, March 3rd. This is huge. If you've been following the other Open Lab releases, they've been getting scarily close at like a tiny sliver of the price of Opus 4.6 and 5.3 Codex, and it looks like this one might actually beat them. The leaked benchmarks not verified, but I would not be surprised if this is true because the other labs have been getting, as I said, very very close. As you can see, greater than 80% on SWE bench verified. It is very likely that this model is better than Opus 4.6 and GPT 5.3 Codex at coding. The leaks say the pricing will be 28 cents per million output tokens. For reference, Opus 4.6 is $25 per million. That's about 90x cheaper. Main flagship models leaked to be a trillion parameters, up from about 600 billion for V3. And allegedly there's going to be a V4 lite model which is 200 billion parameters and code name "sealio-lite". I'm excited for this because I've been spending so much money on tokens for coding. Keep your eye out rumors say this week, some say tomorrow, March 3rd. We will see.