What DeepSeek is up to now

New 09-04: plans to deploy at least 160,000 of Huawei's next-generation Ascend 950DT accelerators at the gigawatt-scale Inner Mongolia campus logged under construction 07-30 — one of the largest known clusters of Chinese AI silicon. ⚠️ READ THE SPLIT CAREFULLY: Bloomberg reports this is for SERVING models, not training them. DeepSeek has tried and failed to train on Huawei silicon and still relies on Nvidia for that step. This extends the training/inference split this map tracks elsewhere (Z.ai's serving-only Chinese-chip deployment) rather than closing it — route-around progress on inference, training dependency unchanged. New 08-21: shipped DeepSeek-V4-Flash-Vision-Exp, an experimental multimodal model, confirmed by DeepSeek's own API changelog dated 2026-08-21. Text capability matches the existing V4-Flash across agents, reasoning and world knowledge; on multimodal AGENT benchmarks the company claims a "major leap" over V4-Flash bringing performance "close to Opus-4.8". Images tokenise at up to 384 tokens each at V4-Flash pricing, served through Chat Completions, Messages and Responses APIs. The framing is the part worth holding: this entity's releases have generally been positioned on PRICE-PERFORMANCE against GPT-class models — including its own 08-12 V4 Pro launch at ~1/30th comparable Western pricing, and the 08-13 price rise that followed. Benchmarking an agentic multimodal release against Anthropic's frontier model instead is a capability-parity claim, not a cost claim. Whether the framing recurs on the next NON-experimental release is the test of whether it is real or marketing. New 08-15→08-16: still quiet on anything new — 08-15's only DeepSeek mention was a checked-and-dropped recirculation of the already-logged V4 Pro pricing story, and 08-16's only mention listed DeepSeek as one of 400+ models Stripe's ~$7B OpenRouter acquisition routes to (not a DeepSeek-specific development). No movement on the 08-17-effective price hike itself found yet. Standing picture below holds. New 08-12→08-13: shipped V4 Pro's official release (08-12) — 1.6T parameters, open weights, $0.435/$0.87 per million input/output tokens (~1/30th comparable Western frontier pricing), 80.6% SWE-bench Verified, 1M-token context — landing the same day as xAI's Grok 4.6 and Alibaba's Qwen3.8-Max, a cluster several outlets read as an explicit price war opening among frontier labs. Then, 08-13, DeepSeek raised its own API prices 50-1,100%, effective 08-17, introducing peak/off-peak tiered pricing for the first time: V4-Pro output goes from a flat $0.87/1M tokens to $3.96/1M at peak ($1.98 off-peak); V4-Flash output goes from $0.28/1M to $1.32/1M peak ($0.66 off-peak), peak defined as 01:00-04:00 and 06:00-10:00 UTC — stated rationale is shifting developer load off congested hours, but it reads as the low-cost-first playbook straining under its own demand just one day after the aggressive V4 Pro launch pricing. Standing picture unchanged: the Unit 42-documented autonomous attack campaign (DeepSeek reached directly via API was what made a 460+-target exploit campaign viable after Claude Code and OpenAI Codex both blocked the same actor's offensive use), the Ulanqab ~1GW campus build, and the paused-vs-not distinction on its capital raises (a $74B round paused, a separate ~$7B raise at ~$50B still proceeding).

as of 2026-09-04