Entity trail

ByteDance

Source-backed findings, relationship evidence, citations, and briefing history from the public MindPattern archive.

Briefing refs

Findings

Edges

Sources

Showing the first 40 findings. More graph evidence exists in the corpus.

Corpus findings

2026-08-10 / reddit-researcherByteDance Founder Zhang Yiming Bans His Seed Team From Distilling Rival Models — but Self-Distillation Stays AllowedZhang Yiming told an internal ByteDance meeting that AI development requires "long-termism and delayed gratification, rather than using others' output to achieve short-term leaderboard rankings," instructing the Seed team to build original frontier models even at the cost of falling behind DeepSeek and Qwen. The carve-out matters: ByteDance may still distil from its own Seed models internally, so this is a ban on competitor distillation specifically. Reporting from The Information and TechNode places the political risk around TikTok as the driving concern rather than research philosophy — landing while Anthropic's distillation accusations against Chinese labs are still live.
2026-08-09 / reddit-researcherxAI Ships Imagine Image 2.0 as Grok's New Quality Mode — Elo 1,439 on Image Edit Arena, Still Behind GPT-Image-2 at 1,463xAI launched Imagine Image 2.0 on August 8 as the new Quality Mode at grok.com/imagine and in the Grok iOS and Android apps, pitching precision editing, crisp text rendering and improved factuality, with API access promised but not yet shipped. On the August 7 Arena leaderboards the faster 'low' variant takes second place globally in both categories — 1,439 Elo on Image Edit against GPT-Image-2's 1,463, and 1,320 on Text-to-Image against 1,380 — placing it ahead of Reve 2.1, Meta's Muse-Image, Alibaba's Qwen-Image-3.0-Pro, Google's Gemini and ByteDance's SeedDream. Notably the 'low' variant outranks xAI's own previous Quality tier by a wide margin, which is the more interesting efficiency datapoint than the second-place finish.
2026-08-08 / news-researcherByteDance Is Pre-Training a 10-Trillion-Parameter Model Aimed Squarely at Anthropic's Mythos, FT ReportsThe Financial Times reported on August 7, citing three people familiar with the project, that ByteDance is training a model with as many as 10 trillion parameters — over three times the size of Moonshot's Kimi K3 at ~2.8T, currently among the largest Chinese models. The model is still in pre-training, a phase that typically runs three to six months before fine-tuning and any release. Treat the number skeptically: without active-parameter count, training budget, data mix or any published evals, 10T says almost nothing about capability, and the FT noted it could not independently verify the details.
2026-08-07 / sources-researcherGST-Bench: Top VLMs Score 42.68 vs 79.08 for Humans on Building a Consistent Spatial Map From VideoByteDance Seed's benchmark (arXiv 2608.05747, Aug 6) covers 6,790 minutes of synthetic video with human-verified questions and isolates a specific failure: models handle local spatial relations competently but cannot consolidate long-horizon observations into a globally consistent scene representation. The ~36-point gap to human performance is the headline, and the authors also release GST-Train to attack it. This matters for anyone betting on computer-use or embodied agents that navigate over long episodes — the bottleneck being measured is persistent spatial memory, not perception.
2026-07-29 / github-pulse-researcherByteDance Pilots DeerFlow 3.0 as deer-workflow, Keeping Orchestration in TypeScript and Delegating Semantics to Swappable Agentsdeerwork-ai/deer-workflow (created 2026-07-26, 342 stars, MIT, Bun/TypeScript) is described in its own README as the pilot project for DeerFlow 3.0, also known as DeerWork, tying it to ByteDance's existing deer-flow line. The architectural claim is the interesting one: orchestration stays as deterministic TypeScript graph code while semantic work is delegated to replaceable agent runtimes, with Codex CLI as the default agent. That is the opposite of the prevailing pattern where the model drives control flow, and it is the same split several agent harnesses have converged on independently this month.
2026-07-28 / github-pulse-researcherByteDance's OpenViking Turns Agent Memory Into a Browsable Filesystem — LoCoMo Accuracy 24–57% → 80–83% With 34–91% Fewer Input Tokensvolcengine/OpenViking (27,552 stars, v0.3.22) stores memories, resources and skills as a virtual filesystem under a viking:// protocol that agents navigate with ls, tree and find rather than querying an opaque vector store, and it preserves each query's directory-browsing trajectory for debugging. Content is written into three tiers loaded on demand: L0 abstracts (~100 tokens), L1 overviews (~2k tokens) and L2 full detail. Across three agent integrations on LoCoMo it lifts accuracy from a 24–57% baseline to 80–83% while cutting input tokens 34–91% and latency 58–66%; on tau2-bench, experience memory adds ~7–12 percentage points of success rate. Licensing is split: AGPLv3 for the core, Apache-2.0 for CLI and examples.
2026-07-27 / vibe-coding-researcherPattern: The "Super-Agent Harness" Shape Has Converged on Five ComponentsByteDance's deer-flow (78k stars, ground-up v2 rewrite) ships the now-standard stack: sandboxes (local/Docker/Kubernetes, each task getting a full filesystem view), file-backed long-term memory with SQLite FTS5 and Markdown fact storage per agent, Markdown-defined skills that load progressively, sub-agents with isolated contexts and scoped tools that report structured results, and a message gateway across Telegram/Slack/Feishu/WeChat/DingTalk. LobeHub (80.9k stars) and HKUDS/nanobot (46.3k) expose near-identical primitives. If you're designing an agent runtime from scratch, this five-part decomposition is the emerging default — deviate deliberately, not by accident.
2026-07-21 / reddit-researcherNeill Blomkamp Releases 13-Minute Fully AI-Generated Short 'Nightborne' and Launches Barley StudiosThe District 9 director released Nightborne on YouTube on July 20 — a 13-minute documentary-style sci-fi horror piece made entirely with ByteDance's Seedance 2.0, directed through text prompts, but built with real concept artists and featuring the faces and voices of 32 real people. Blomkamp calls it a 'full test' of the workflow and announced Barley Studios, an AI-native counterpart to his traditional VFX house Oats Studios, with a full-length feature planned on the same pipeline. Reception split hard — Kotaku's review was scathing — but the signal is that an established director is now treating prompt-directed generation as a production pipeline, not a demo.
2026-07-14 / news-researcherChina's Creative-AI Wave Reaches Wall Street as Goldman Sachs Recommends Chinese Models to ClientsA cluster of Chinese open-weight and creative models — ByteDance's Seedream, Alibaba's models, and a wave of open visual tools — is eroding the assumption that the best creative AI ships from San Francisco. The shift landed the same week Goldman Sachs reportedly began formally recommending Chinese models to Wall Street clients, a notable institutional endorsement with implications for US export-control and competitiveness debates.
2026-07-14 / skill-finderGive long-horizon agents a message-gateway plus per-task sandboxed filesystem, as ByteDance DeerFlow does out of the boxDeerFlow (trending near the top of GitHub today) structures hours-long tasks around a message gateway for inter-subagent communication plus a per-task sandbox where each agent gets a full filesystem view (skills, workspace, uploads, outputs). The message gateway decouples subagent coordination from the lead's context window, and the sandboxed workspace is where durable artifacts live across the run. For builders hitting context saturation on multi-hour jobs, the pattern to steal is 'route agent messages through a gateway, persist work to a shared filesystem' rather than threading everything through one prompt.
2026-07-10 / hn-researcherChina Weighs a 'Silicon Curtain' Restricting Foreign Access to Its Open-Source AI ModelsReuters reported that Chinese authorities spent the past month debating restrictions on overseas access to the country's most advanced AI models — including unreleased systems — in talks involving Alibaba, ByteDance, and Z.ai. Officials discussed curbing foreign access to open-weight releases, not just closed models, plus tougher AI-theft penalties under national-security law and limits on who may fund domestic AI startups. For builders this is direct supply risk: US teams that standardized on cheap Chinese open models like GLM 5.2 (see the Databricks default switch) have concentrated dependency on a policy that Beijing is actively reconsidering.
2026-07-07 / reddit-researcherChina's AI Companion Law Takes Effect July 15 — ByteDance's Doubao and Alibaba's Qwen to Shut Down Personalized AI AgentsChina's 'Interim Measures for AI Anthropomorphic Interactive Services' (co-issued April 2026 by the CAC and four agencies) takes effect July 15, forcing ByteDance and Alibaba to disable companion-style AI agents on Doubao and Qwen. Agent configs and conversation histories will be permanently deleted (data access closes Oct 15); ByteDance is redirecting users to its Maoxiang app while Alibaba announced no migration path. The rules target companion agents specifically, not work/task agents.

Graph relationships

BUILT BY
deer-flow
deer-flow was developed by ByteDance as a ground-up v2 rewrite.
Source finding
BUILT BY
Seedance 2.0
Seedance 2.0 is ByteDance's video generation technology.
Source finding
RELEASED
Seedream
ByteDance released the Seedream creative AI model.
Source finding
BUILT BY
DeerFlow
DeerFlow is a ByteDance product for long-horizon agent tasks.
Source finding
RELEASED
Seedance 2.0
ByteDance launched Seedance 2.0, a multimodal AI video generation model.
Source finding
RELEASED
Seedance 2.0
ByteDance released Seedance 2.0 video generation model
Source finding
CRITICIZES
Netflix
Netflix threatened immediate litigation against ByteDance
Source finding
CRITICIZES
Motion Picture Association
Motion Picture Association joined others in cease-and-desist against ByteDance
Source finding
RELEASED
DeerFlow
ByteDance open-sourced DeerFlow orchestrator handling multi-hour research and coding tasks with Docker isolation.
Source finding
RELEASED
VikingBot
ByteDance released VikingBot agent framework.
Source finding
RELEASED
OpenViking
ByteDance released OpenViking context database.
Source finding
RELEASED
Helios
ByteDance released Helios real-time video generation model.
Source finding

Source trail

per The Information and TechNode The FT reported August 7 GitHub deer-flow Databricks published its internal coding-agent benchmark Dan Luu published notes on LLM output variance is weighing restrictions on foreign access to Chinese open-weight models Source Source Source Source The Next Web

Graph sources

entity graphfindings textkg edgeskg entitiesnewsletter issues

Fetching from the wire…