Claude Code sessions can now message each other, and it kills the copy-paste relay
On August 7, Anthropic shipped cross-session messaging in Claude Code v2.1.224: one running CLI session can discover your other sessions with a ListAgents tool and deliver a short text message with a SendMessage tool. The receiving session reads it between tool calls, or starts a fresh turn if it was idle. No setup, no config file, no server — if you are on v2.1.224 on macOS or Linux, it is just on.
What actually travels is deliberately narrow: plain text only, no files, no conversation history, no context. If you want another session to inherit full context, the docs point you to resume instead. The design that matters is the permission behavior. A session in bypass mode does not get to whisper straight into another bypass-mode session — that pair holds messages for your approval, with a five-minute expiry on the dialog. Admins can refuse inbound messages or disable SendMessage/ListAgents org-wide. Loops are throttled, identical repeats are dropped, and an inbox caps at 50 messages.
I have been doing the copy-paste-between-terminals dance for two years, and the four documented use cases — hand over a finding, coordinate worktrees, check on a long-running job, reply from another machine — are exactly the ones that cost me real time. This is the same philosophy Anthropic applied to auto mode: capability grows, and the safety moves into the channel itself rather than into a rubber-stamp dialog. The honest caveat is that delivery is not guaranteed, and a cross-machine session can reply but never initiate. Still, for parallel-agent workflows, this is the first first-party fix for what was previously a human relay problem.
— Anthropic · ClaudeDevs
🔗 Anthropic docs · Anthropic (X)
ByteDance is reportedly pre-training a 10-trillion-parameter model
The Financial Times reported on August 7, citing three people with knowledge of the matter, that ByteDance is pre-training an AI model with as many as 10 trillion parameters — more than three times the size of Moonshot's Kimi K3 (2.8T) and close to industry estimates for Anthropic's Mythos 5 (~8T). Pre-training of this kind typically runs three to six months, and the final size is not locked until late in the process. Reuters could not independently verify the report, and ByteDance did not respond to requests for comment.
The context makes this less surprising than it sounds. 晚点 LatePost had already reported ByteDance was discussing a model above 5T parameters, led by Seed Foundation head 项亮 (Xiang Liang) in collaboration with pre-training data lead 沈科 (Shen Ke). At an all-hands meeting on August 6, CEO 梁汝波 admitted Doubao's AI coding is not a strength and cited Anthropic's Claude Code as the benchmark, while founder 张一鸣 has reportedly pushed the team to chase "world-class model capability" rather than short-term wins. ByteDance also told its team not to use distillation from other companies' models, which analysts say has slowed its progress while keeping its training independent.
I read this as a genuine inflection for China's model race, not just another headline number. ByteDance is the one Chinese lab that stays fully closed-weight — Doubao has 324 million monthly users and Seedance is already a leading video generator — and a 10T model would put it in a class no Chinese lab has shipped. The parameter count itself is a weak proxy for capability, and the FT report is a leak, not a release. But the strategic direction is unambiguous: after a year of being seen as behind on frontier intelligence, ByteDance is betting its compute budget on scale, and the market will find out in 2026 H2 whether the bet pays off.
— Financial Times · 智东西
🔗 Financial Times (via Reuters) · 智东西
Ant Group open-sources Ling-3.0-flash, a 124B model that activates only 5.1B
Ant Group's Bailing (百灵) team released Ling-3.0-flash as open weights on August 7 under a permissive license, with weights on Hugging Face and ModelScope. The architecture is a native hybrid-linear MoE: 124B total parameters but only 5.1B activated per token, alternating KDA and MLA layers at a 5:1 ratio, with a 256K context window that scales to 1M. The model was refined across more than 10,000 interactive agent environments, and Ant positions it as the "execution node" in a planning-execution split — deep planning stays on ultra-large reasoning models, and Ling-3.0-flash handles fast, cost-controlled execution.
The numbers back the positioning. On the Artificial Analysis Intelligence Index, Ling-3.0-flash runs about $0.04 per task with a 1.4-minute average decode time, sitting in the "intelligence-cost" and "intelligence-latency" advantage zones. It hits 353 tokens/s output on AA's charts, and Ant claims over 1,100 tokens/s in a high-performance configuration. Quantized FP4/INT4 builds run end-to-end on a single NVIDIA DGX Spark, which matters for enterprises that cannot let data leave the building. Huawei's Ascend stack added 0-day support, and Ant says the model cuts time-to-first-token on long inputs by 60% to over 80% via hierarchical caching.
This is the least flashy release of the week and one of the most practical. Ant is not trying to out-parameterize anyone; it is shipping a small-active, big-knowledge model aimed at production agent loops, where latency and per-call cost decide whether a workflow survives. The trade-off is real — 5.1B active parameters means you are not getting frontier multi-step reasoning on hard problems — but as an execution tier under a planning model, the economics are hard to argue with. The open license and the single-DGX-Spark deployment story give it a clear niche: enterprises that want agent workloads on their own hardware without paying frontier API prices.
— Ant Group · Hugging Face
🔗 Ant Group (Business Wire) · Hugging Face
MiniMax H3 tops Design Arena in three video categories and Hugging Face trending
MiniMax's open-source video model H3 has had a week. On August 6, the crowd-sourced Design Arena benchmark put it first in multi-image-to-video, image-to-video, and video editing — ahead of closed models like Seedance 2.0, Grok Imagine Video 1.5, and Gemini Omni Flash. Within three days of release it hit the top of Hugging Face's trending chart, overtaking DeepSeek V4 Flash, and more than 100 partners did Day-0 integration across chips and inference frameworks. Stable Diffusion founder Emad Mostaque posted "Bravo to MiniMax," and a16z's Justine Moore said H3 was the first model to pass one of her challenges.
The technical core is Context-IR: the model ingests text, images, video, and audio, and compresses a raw input that might cost 100K tokens into an average of about 4,000 tokens of structured context — while preserving which character appears when, which audio belongs to which frame, and what the user actually wants changed. That unified context turns text-to-video, motion reference, character replacement, and video editing into different commands within one system rather than separate products. The stock market noticed: MiniMax shares rose about 25% in four trading days after the July 31 release, and Jefferies reiterated a Buy with a HK$1,118 target.
I think the important story here is not the benchmark top — benchmarks in this space decay fast — but that a Chinese open-weight model beat several flagship closed video models on a crowd benchmark while the market re-rated the company in real time. H3's Day-0 ecosystem response (chips, inference frameworks, 100+ partners) is the part that compounds, because it is the same playbook DeepSeek ran on the text side, applied to video. The open question is monetization: H3 is free weights, and MiniMax's business model still leans on API and paid tiers. For now, though, this is the strongest signal yet that the "DeepSeek moment" is spreading to generative video.
— Design Arena · 澎湃新闻
🔗 Design Arena · 澎湃新闻
DeepSeek quietly restarts its second funding round at a ¥500B pre-money valuation
Caijing reported on August 5 that DeepSeek has restarted its second external funding round after a brief pause, planning to raise ¥50 billion (about $7.4 billion) at a pre-money valuation of roughly ¥500 billion (~$74 billion) — about 43% above the post-money value of its first round two months ago. Signing is targeted for late August. The round originally opened in mid-July and was paused at month's end, reportedly because founder 梁文锋 was unhappy with widely circulated content based on a leaked investor-meeting transcript. Some waitlist investors say they have not yet been notified of the restart, so outreach remains selective.
The deal math is bracing. The first round, which closed in June, also raised ¥50 billion at a valuation above ¥350 billion — the largest first raise in Chinese AI history — with investors including the National AI Industry Investment Fund, Tencent, CATL, NetEase, JD.com, and IDG Capital. Interest in the first round reportedly exceeded ¥100 billion, so at least ¥50 billion of demand was left waiting at the door. If the second round closes, DeepSeek will have raised over ¥100 billion in two rounds within a few months, far outpacing rivals. The backdrop is V4-Flash's public beta (July 31), which scores 50 on the Artificial Analysis Intelligence Index — second among domestic models — while pricing output at $0.28 per million tokens and topping OpenRouter's weekly token rankings.
One deal participant quoted by Caijing put it bluntly: pricing an LLM company "is essentially an options trade, not a cash-flow-based financial model." I think that is the right frame for the whole Chinese tier right now — Moonshot went from $18B to $50B in three months, and DeepSeek's ¥500B pre-money follows the same logic, where the option is on frontier capability rather than today's revenue. The risk is the same one that hit the sector in 2025: valuations that are entirely hostage to the next model release. If V4's next iteration stumbles, the paper mark-up has a long way to fall.
— 财经 (Caijing) · Yicai Global
🔗 Caijing (via Yicai) · 香港文匯報
H200 shipments to China are near zero, and domestic chipmakers are eating the budget
A U.S. Commerce Department official confirmed at a congressional hearing on August 8 that, despite export licenses being issued for NVIDIA's H200, actual shipments to China are negligible — effectively zero. That is a stark reversal of Jensen Huang's March framing of large China H200 orders and production restarts, and it confirms what CNBC had reported: NVIDIA's share of the Chinese AI chip market fell from roughly 95% in 2023 to near zero for new H200 shipments by mid-2026. In June, BIS also closed the "subsidiary loophole," requiring licenses for advanced chips sold to any China-headquartered company anywhere in the world.
The money is moving. A July survey of 60 Chinese tech executives found they expect to put 46% of their AI accelerator budget into domestic chips over the next 12 months, up from 30% today. Cambricon (寒武纪) reported H1 revenue near ¥6 billion, up over 108% year on year with net profit up 122.61%; Moore Threads guided H1 revenue of ¥1.65–1.75 billion, up 135–149%; brokerages forecast DaysiZhixin at about ¥3.04 billion, nearly tripling. On top of that, Beijing is planning roughly ¥2 trillion over five years for national data centers, with over 80% of core hardware required to come from domestic suppliers. Industry projections put China's AI chip self-sufficiency at 70% by 2029, up from 42% in 2025.
I do not think this is a clean "decoupling is working" story. What the numbers show is a policy loop that feeds itself: export controls stay restrictive, so budgets shift to domestic silicon, which validates the controls, which justifies more budget shift. The interesting risk is for NVIDIA's long game — Huang keeps warning that inconsistent export rules push Chinese firms toward self-sufficient supply chains, and the hearing testimony is the strongest evidence yet that the shift is already priced into procurement. For the domestic vendors, the windfall is real, but it is also a subsidy-driven boom; the question is whether Cambricon and Moore Threads can hold margins once the low-hanging substitution demand is met.
— CNBC · 新浪财经
🔗 CNBC (via 163/IT之家) · The Index Today
Unitree prices its STAR Market IPO at ¥150.80 — and DeepSeek and Tencent are on the strategic list
Unitree (宇树科技), the "first humanoid robot stock" on the STAR Market, priced its IPO on August 6 at ¥150.80 per share, raising roughly ¥6.1 billion — well above the originally planned ¥4.2 billion, so the deal was oversubscribed. The issue implies a market cap around ¥61 billion at a P/E of about 219 times. Nine institutions secured strategic placement, and the list is notable: DeepSeek and Tencent are both among them. At the August 7 online roadshow, founder 王兴兴 said the embodied-AI industry is at the equivalent of the early PC stage, arguing the sector is far from maturity.
The fundamentals behind the hype are real but concentrated. Unitree shipped over 5,500 humanoid robots in 2025, roughly 32.4% global share and first in the world; its four-legged robots have cumulative shipments above 33,000, close to 60% global share, and it is one of the few robot makers that is actually profitable. But the "strong at the body, weak at the brain" critique follows it — analysts note its edge is motion control, not the world-model or VLA intelligence layer — and Unitree plans to direct close to half the IPO proceeds into model R&D. That is a direct acknowledgment that the valuation ceiling depends on closing the cognitive gap, not shipping more units.
I think this IPO is the real test of how the market prices embodied AI, and the 219x P/E says the market is buying the story rather than the cash flows. The strategic placement is the detail worth sitting on: DeepSeek investing in a robot maker, and Tencent alongside, is capital lining up behind the "robots need frontier brains" thesis — the same logic behind NVIDIA's robotics push and the 智元 HK IPO that followed the same week. Whether ¥61 billion is sane depends entirely on whether the brain gap closes in the next two years. If it does, this looks early; if not, 219x has a long way to compress.
Top comments (0)