← Back to feed

ByteDance’s volume-tier model got new weights. The seed-2-1-turbo-20260810 build went live on August 10, the first Turbo refresh since the Seed 2.1 family launched on June 23. Three days later, DeepSeek V4 Pro shipped its official GA build (DeepSeek-V4-Pro-0813), ending a four-month preview period.

The convergence is not coincidence — it is the same market. Seed 2.1 Turbo and DeepSeek V4 Pro are the two main Chinese-developed options for cheap frontier-adjacent reasoning. They landed in the same news cycle with fundamentally different product bets.

The Two Models

Seed 2.1 Turbo (ByteDance, seed-2-1-turbo-20260810)

• $0.50 input / $2.50 output per 1M tokens
• 262K context
• Multimodal: text, image, video input in a single call
• Hosted API only; one provider (ByteDance Seed)
• Terminal-Bench 2.1: 67.6%
• NL2Repo: 43.7%
• Workspace Bench: 54.7%
• CyberGym: 67.0%
• MobileWorld: #1 (mobile GUI task execution)
• BeyondAIME: 88.0%

DeepSeek V4 Pro GA (DeepSeek-V4-Pro-0813)

• $0.44 input / $0.87 output per 1M tokens
• Open weight, MIT licensed — run on your own GPUs
• Text only
• Terminal-Bench 2.1: 87.9% (vendor-reported)
• DeepSWE: 62.7% (vendor-reported; 5x jump from preview)
• SWE-bench Verified: 96.4% (independent, vals.ai neutral bash harness)

Where They Actually Compete

The vendor benchmarks diverge enough to make head-to-head comparison unreliable. DeepSeek’s official figures (Terminal-Bench 87.9, DeepSWE 62.7) are from its own harness against its own task set. ByteDance’s figures (Terminal-Bench 67.6, Workspace Bench 54.7) come from the June launch and were not independently re-verified for the August build.

The third-party picture cuts differently. On BenchmarkList’s cross-model runs — the two long-horizon coding benchmarks where both models have external numbers — Seed 2.1 Turbo leads the DeepSeek V4 Pro Max variant by 36.8 points on NL2Repo and 14.3 points on CyberGym. DeepSeek’s spread between its official agentic claims and third-party reads remains wide enough that production buyers should factor it in.

The Fork in Strategy

DeepSeek V4 Pro’s case is cost and openness: $0.87 per million output tokens, MIT license, self-hostable, text-only. If you are running pure-text LLM inference and want the option of on-prem deployment, there is no cheaper frontier-adjacent option with a comparable SWE-bench number.

Seed 2.1 Turbo’s case is the workload type. It is the cheapest frontier-adjacent model that accepts video input in the same API call as text and images, and ByteDance is actively maintaining it — the August 10 build is the signal. The model targets agent workflows that need visual context: document parsing, GUI interaction, mobile automation. MobileWorld #1 is not a random benchmark choice.

The $2.50 output price is above DeepSeek’s $0.87, but Seed 2.1 Turbo is not competing on raw token cost. It is competing on being the only maintained, multimodal, deep-reasoning option in the Chinese-developed tier that a developer does not have to host themselves.

ByteDance launched the Seed 2.1 family on June 23 with a Pro and Turbo split. The August 10 Turbo rebuild — seven weeks into the product’s life — suggests the update cadence will be faster than the typical frontier lab cycle. Whether the next refresh narrows the Terminal-Bench gap with DeepSeek V4 Pro is the number to watch.