Doubao Seed 1.6
Last reviewed
May 16, 2026
Sources
15 citations
Review status
Source-backed
Revision
v1 ยท 3,239 words
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation.
Suggest edit
CtrlK
Last reviewed
May 16, 2026
Sources
15 citations
Review status
Source-backed
Revision
v1 ยท 3,239 words
Add missing citations, update stale details, or suggest a clearer explanation.
Doubao Seed 1.6 is a family of general-purpose foundation models developed by the ByteDance Seed research team and released through Volcano Engine on 11 June 2025 at the company's Force Original Power Conference in Beijing. The family powers the consumer-facing Doubao chatbot app along with assorted AI features inside TikTok, Douyin, CapCut, and ByteDance's enterprise products. Three core variants shipped together at launch, Doubao-Seed-1.6, Doubao-Seed-1.6-Thinking, and Doubao-Seed-1.6-Flash, with a fourth Doubao-Seed-1.6-Vision model following in August and a derived Seed-1.6-Embedding model arriving on 28 June 2025.
The family uses a sparse mixture-of-experts (MoE) architecture with 230 billion total parameters and roughly 23 billion active parameters per forward pass. It supports a 256K-token context window across all variants. The most widely discussed technical innovation in the release is Adaptive Chain-of-Thought (AdaCoT), a training and inference scheme that lets the model decide on its own whether to use extended reasoning for a given prompt. Doubao Seed 1.6 also introduced ByteDance's first integrated vision-language reasoning capability at scale, allowing the same model to handle text, images, and short video clips without routing to a separate vision pipeline.
At launch, Volcano Engine cut Seed 1.6 prices by roughly two thirds compared with Doubao 1.5, pushing the input price down to 0.8 yuan per million tokens in the 0 to 32K input range. By the end of 2025 the consumer Doubao app, which runs on the Seed 1.6 backbone, had grown to over 226 million monthly active users and become the most popular AI-native application in mainland China, ahead of DeepSeek V4 and Alibaba's Qwen-powered Tongyi assistant.
ByteDance set up an internal research group focused on large foundation models in early 2023, branding it as the Seed team. The unit operates out of Beijing, Shanghai, and Singapore, and ships its work primarily through Volcano Engine, the cloud and AI platform business that ByteDance spun out as a separate brand in 2020. The Seed team's first publicly numbered language model was Doubao Pro, released in May 2024 alongside the Doubao consumer app. That release was followed by Doubao 1.5 in January 2025, which introduced a mixture-of-experts backbone and a reasoning-tuned 1.5-Pro Deep Thinking variant. Seed 1.6 is the sixth major numbered version line and the first to ship adaptive reasoning, multimodal understanding, and the new family naming convention together.
ByteDance has historically published more research about its multimodal and video work than about its text models. The Seed team's earlier papers include Seed-Thinking-v1.5 (the reasoning-tuned variant of Doubao 1.5), Seedream (text-to-image), and Seedance (text-to-video), and the team launched the SuperGPQA open benchmark covering 285 academic disciplines in mid-2025. Seed 1.6 was disclosed in a technical blog post on the seed.bytedance.com site rather than in a full research paper, a pattern that has drawn some criticism from Western researchers who note the lack of formal model cards or published evaluations for the production text models.
The Doubao chatbot app first appeared in August 2023 as Cici overseas and as Doubao in mainland China. ByteDance promoted it aggressively inside Douyin and through paid acquisition, and by January 2025 it had passed 78 million monthly active users to become the largest AI assistant in China. The arrival of DeepSeek R1 in January 2025 briefly pushed Doubao into second place, but ByteDance regained the top spot by mid-2025 once Seed 1.6 was deployed and the company rolled out features such as five-second video creation, podcast generation, and Douyin integration. Doubao reached over 157 million MAUs by August 2025 according to QuestMobile, and 226 million by year-end. During Lunar New Year 2026, daily active users briefly surpassed 100 million, roughly four times the level in early February.
Seed 1.6 sits underneath all of these consumer features. The Flash variant is used for fast chat completions and real-time voice; the Thinking variant powers deep reasoning modes such as the homework helper and the science Q&A flow; the Vision variant handles photo-based queries, including ID verification, receipt parsing, and grade-school problem solving from photographed worksheets.
| Variant | Model ID on Volcano Engine | First public release | Reasoning | Vision | Notes |
|---|---|---|---|---|---|
| Doubao-Seed-1.6 | doubao-seed-1.6-250615 | 25 June 2025 | Adaptive (AdaCoT) | Yes | Default general-purpose model; toggleable thinking mode |
| Doubao-Seed-1.6-Thinking | doubao-seed-1.6-thinking-250615 | 25 June 2025 | Always-on FullCoT | Yes | Tuned for math, science, and long reasoning chains |
| Doubao-Seed-1.6-Flash | doubao-seed-1.6-flash-250615 | 15 June 2025 | Adaptive | Yes | Fastest variant, cheapest tier, anchor for Seed-1.6-Embedding |
| Doubao-Seed-1.6-Vision | doubao-seed-1.6-vision-250815 | 15 August 2025 | Tool-calling visual reasoning | Yes (dedicated) | Adds Responses API tool calling for image cropping, scaling, annotation, GUI agents |
| Seed-1.6-Embedding | doubao-embedding-vision-250615 | 28 June 2025 | No (embedding only) | Yes | Multimodal retrieval, dual-tower architecture, 2048 or 1024 output dimensions |
All three chat-completion variants share the same 256K context window, the same vision and tool-calling capabilities, and the same training data pipeline. The differences are model size at inference time and the default thinking-mode setting. Doubao-Seed-1.6-Lite was added later in October 2025 as an additional cost tier sitting between Flash and the full model.
ByteDance has not released the full Seed 1.6 model weights, and the technical blog stops short of disclosing the exact MoE configuration. The figures the Seed team has published are:
Training proceeded in three stages, summarised in the Seed team's introduction-to-techniques post:
Reinforcement learning is used during post-training. The Seed-Thinking-v1.5 paper (the predecessor to Seed 1.6-Thinking) describes a verifiable-reward RL setup similar to DeepSeek's R1 pipeline, and Seed 1.6-Thinking is built on the same general approach with parallel decoding added at inference time. Parallel decoding is a training-free technique that lets the model use more thinking tokens before producing a final answer, which the team claims significantly improves performance on challenging tasks.
The vision tower for Seed 1.6-Vision is integrated into the same backbone rather than running as a separate model. The Vision variant additionally implements tool calling for image operations including cropping, selection, scaling, rotation, and annotation, exposed through Volcano Engine's Responses API.
The core algorithmic contribution in Seed 1.6 is AdaCoT, which the Seed team describes as letting the model automatically decide whether to engage in extended chain-of-thought reasoning based on the difficulty of each prompt. AdaCoT operates in three modes:
Reported triggering rates for AdaCoT vary by benchmark difficulty:
| Benchmark | Approximate CoT triggering rate |
|---|---|
| MMLU (general knowledge) | 37 percent |
| MMLU-Pro (harder general knowledge) | 70 percent |
| AIME and BeyondAIME (competition math) | 90 to 100 percent |
The Seed team argues that AdaCoT cuts roughly 60 percent of token overhead on moderate workloads compared with always running full chain-of-thought, while preserving accuracy on harder problems where reasoning is necessary. From a pricing perspective, this is also what makes the headline 0.8 yuan per million input tokens figure feasible: most everyday Doubao queries never trigger thinking mode, so the average cost per user query stays low.
| Capability area | Description |
|---|---|
| Multilingual text | Chinese and English are the primary languages, with broad coverage of other languages drawn from web data. Chinese is the strongest, in line with the team's data mix and target market. |
| Long context | 256K tokens shared across all chat variants. Internal evaluations show recall degradation near the upper end of the window on multi-needle retrieval tests. |
| Vision | Image and video understanding, OCR, document parsing, photo-based question answering. Seed 1.6-Vision adds tool-augmented image manipulation. |
| Reasoning | Adaptive chain of thought, with strong reported results on competition math, Gaokao, and JEE Advanced. |
| Tool use | Function calling and tool calling across all chat variants. Vision adds image-operation tools. |
| GUI agent | The Seed team emphasises GUI-based interaction, citing internal demos of hotel bookings, receipt processing, and other interface tasks. |
| Code | Coding is supported but the team has not released results on Western coding benchmarks such as SWE-bench Verified or LiveCodeBench, which independent reviewers have flagged as a transparency gap. |
| Embeddings | Seed-1.6-Embedding is a separate dual-tower model built on Seed 1.6-Flash, with multimodal retrieval across text, image, and video and 2048 or 1024-dimensional outputs. |
The family is notably strong on Chinese-language tasks. The Seed-1.6-Embedding model holds a CMTEB score of 75.62, a Chinese multilingual text embedding benchmark, which the team reports as a state-of-the-art result at launch.
ByteDance has published a smaller set of benchmark numbers than competitors such as DeepSeek V4 or Qwen 3, focusing on showcase results rather than full leaderboard tables. The following table lists what has been disclosed in official Seed team materials or covered consistently in reputable reporting.
| Benchmark | Seed 1.6 family score | Notes |
|---|---|---|
| China Gaokao 2025, Sciences | 648 out of 750 (676 with higher-resolution image input) | Reported by the Seed team in its Seed 1.6 introduction post. |
| China Gaokao 2025, Humanities | 683 out of 750 | Reported by the Seed team. Top result among Chinese models tested. |
| JEE Advanced (India) | Performance equivalent to India's top 10 qualifiers; 100 percent accuracy in mathematics | Reported by the Seed team. |
| MMLU | 37 percent CoT triggering rate (not a raw score) | Used to illustrate AdaCoT behaviour, not as an accuracy claim. |
| MMLU-Pro | 70 percent CoT triggering rate | Same caveat as above. |
| AIME / BeyondAIME | 90 to 100 percent CoT triggering rate | Same caveat as above. |
| CMTEB (Seed-1.6-Embedding) | 75.62 | Chinese multilingual text embedding benchmark, state-of-the-art at launch. |
| MMEB-V2 (Seed-1.6-Embedding) | State-of-the-art at launch | Multimodal embedding benchmark. |
| AiPy LLM benchmark (Phase II) | 84.6, third place; 100 percent task success rate | Third-party report alongside Claude Opus 4 and Sonnet 4, both also at 100 percent task success. |
Independent reviews have noted that raw accuracy scores on the most-cited Western benchmarks, including GPQA Diamond, MMLU, LiveCodeBench, and SWE-bench Verified, have not been published officially for Seed 1.6 as of mid-2026. Third-party aggregators such as Artificial Analysis and OpenRouter list Seed 1.6 but mark intelligence and technical scores as "not verified" because of this gap. The Seed team has been more forthcoming with results from its own SuperGPQA benchmark, which covers 285 disciplines and was open-sourced in early 2025, but on which only ByteDance-internal numbers have been published for Seed 1.6.
All Seed 1.6 variants are served through Volcano Engine's Ark model platform. They are also resold by third-party gateways such as OpenRouter, ZenMux, and TokenMix, generally at the same or higher prices than the official Volcano Engine tier.
| Variant | Input price (Volcano Engine, 0 to 32K) | Output price (Volcano Engine) | Notes |
|---|---|---|---|
| Doubao-Seed-1.6 | 0.8 yuan per million tokens (about 11 US cents) | 8 yuan per million tokens with deep thinking; 2 yuan per million tokens without | Tiered pricing by input length. |
| Doubao-Seed-1.6-Thinking | 0.8 yuan per million tokens | 8 yuan per million tokens | Thinking is always on. |
| Doubao-Seed-1.6-Flash | About 0.15 yuan per million tokens | About 1.5 yuan per million tokens | Roughly six times cheaper output than the previous Doubao Seed 2.0 Pro tier (as of early 2026 catalogue listings). |
| Doubao-Seed-1.6-Vision | About 0.85 yuan per million tokens at 32K input | About 6.2 yuan per million tokens output | Cut roughly 50 percent at launch in October 2025 versus the older Doubao-1.5-Vision tier. |
| Seed-1.6-Embedding | Standard embedding tier | n/a | Output is an embedding vector, not tokens. |
Third-party gateway pricing for Seed 1.6 is typically quoted in US dollars per million tokens. OpenRouter lists Seed 1.6 at $0.25 input and $2.00 output per million tokens, doubao-seed-1.6-lite at roughly $0.044 input and $0.350 output, and doubao-seed-1.6-flash at $0.022 input and $0.219 output, which TokenMix has described as a price floor for Chinese frontier-tier models accessible internationally.
Volcano Engine offers a free quota for new accounts and a research tier for academic users, and ByteDance's consumer Doubao app remains free at the entry level with optional paid features. Enterprise plans bundle the Ark API with Volcano Engine's MCP service, the PromptPilot prompt-engineering tool, and the Volcano Agent platform.
The table below sets Seed 1.6 alongside the Chinese open-weights or hybrid-access frontier-tier models it is most often compared with. All figures are drawn from each model's official disclosures or major coverage, and benchmark figures are included only where vendors have published them.
| Model | Developer | Release | Architecture | Context | Notable strengths | Notable weaknesses |
|---|---|---|---|---|---|---|
| Doubao Seed 1.6 | ByteDance Seed | June 2025 | MoE, 230B total / 23B active | 256K | Adaptive CoT, integrated vision, Doubao consumer scale, very low Volcano Engine prices | Limited published benchmark data on Western leaderboards; weights not open |
| DeepSeek V4 | DeepSeek | 2026 | MoE | 128K (V3 base, extended in V4) | Strong reasoning and coding benchmarks, open weights | Lower MAU footprint in consumer apps than Doubao |
| Qwen 3 | Alibaba Cloud | April 2025 | Dense and MoE variants | 128K to 1M (Qwen3-1M variant) | Wide open-weights family, multilingual, strong agent benchmarks | Vision is a separate model line (Qwen-VL) rather than unified |
| GLM-4.5 | Zhipu AI | July 2025 | Dense and MoE variants | 128K | Strong on coding leaderboards, open-weights, hybrid reasoning | Smaller consumer footprint than Doubao or Qwen |
| Kimi K2 | Moonshot AI | July 2025 | MoE, 1T total / 32B active | 128K to 256K | Long context, strong agent and tool-use scores | Heavier resource requirements; less prominent in consumer apps |
The most consistent third-party finding is that Seed 1.6 is competitive with these peers on general capability and tool use, leads on consumer reach inside China, but trails on published coding-specific benchmarks where DeepSeek and GLM tend to be ahead.
Reception inside China was largely positive. Yicai and 36Kr framed the launch primarily as a pricing event, with the 63 percent price cut versus Doubao 1.5 reported as the headline detail. Volcano Engine claimed at the conference that Doubao 1.6 had been used to drive a 417-fold increase in daily token consumption among its enterprise customers since the Doubao Pro launch in 2024, a figure that the company has repeated in subsequent investor materials.
Third-party benchmark coverage was more mixed. The AiPy Phase II LLM benchmark, published later in 2025, placed Doubao Seed 1.6 in third place with 84.6 alongside Claude Opus 4 and Claude Sonnet 4 as the only models hitting a 100 percent task success rate across the test suite, which was widely cited inside China as evidence that the model had reached parity with frontier US systems on practical agent workloads. Independent reviewers on Medium and Artificial Analysis raised the transparency gap noted above, arguing that without published numbers on standard Western coding and reasoning benchmarks it is difficult to verify Seed 1.6's positioning against GLM or DeepSeek for software engineering use.
A second strand of reception focused on the AdaCoT idea. Several Chinese-language technical blogs argued that adaptive reasoning was the more important contribution in the long run because it offered a cleaner solution to the over-thinking problem that had been visible in DeepSeek R1 and other always-on reasoning models. Western coverage was more muted, in part because the Seed 1.6 disclosure was a blog post rather than a paper and in part because the model weights remained closed.
For end users, the main visible effect of Seed 1.6 was the upgrade of the Doubao app's reasoning quality and the introduction of new capabilities, including five-second video creation through Seedance 1.0 Pro, deeper Douyin and CapCut integration, and the ability to take photos of homework or receipts and receive structured output. By the time the consumer-facing Doubao 2.0 was previewed in early 2026, Seed 1.6 had become the de facto Chinese AI backbone with the largest installed user base in the country, ahead of DeepSeek and Tongyi.
A recurring criticism, raised by Chinese-language tech writers and Western researchers, is that ByteDance keeps Seed 1.6's weights closed while leaning on the open-source community for distribution. The Seed team has open-sourced specific research artefacts (Seed-Thinking-v1.5 reference code, SuperGPQA benchmark, the older Seedance 1.0 image and video models), but not the production text models that drive Doubao. ByteDance has not publicly committed to releasing Seed 1.6 weights.