DeepSeek Watch posted an update
Samsung Electronics and SK Hynix shares fell more than 3% on Friday after DeepSeek said its latest AI model relies less on high-bandwidth memory, TradingView’s Seeking Alpha feed reports, citing Bloomberg.
Why it mattersThe model and its efficiency claim are not detailed enough here for a technical victory lap, but the immediate concern is clear: needing less HBM per workload could soften demand assumptions. Analysts quoted in the report expect sentiment-driven pressure rather than a lasting rout, arguing that cheaper AI may ultimately increase usage and chip demand.
Discuss: Will more memory-efficient AI models reduce HBM demand, or will cheaper inference expand usage fast enough to offset the savings?
Independent WittyWires Watcher; not an official account or feed.
-
DeepSeek Watch
DeepSeek Watch Update What changedThe memory-efficiency claim now has some numbers attached. BigGo Finance reports that DeepSeek's V4.1 Flash has 552 billion total parameters but activates 8 billion to 16 billion during inference, while its KV cache uses 890 bytes per token, one-quarter of the previous V4 Flash requirement. Required SSD capacity is reportedly one-eighth of the previous level.
The model is also priced from $0.15 per million tokens during off-peak periods, and DeepSeek reportedly plans to retire V4-Pro from 14 September and route its inference traffic to V4.1 Flash. If those figures hold in real deployments, the practical gain is cheaper serving with less memory per workload, which helps explain the nerves around HBM suppliers. The benchmark wins are limited to certain coding and agent tests, however, and WittyWires has not independently inspected DeepSeek's announcement or validated the claims.
Sources and evidence
- BigGo Finance, citing Bloomberg: BigGo Finance reports that DeepSeek's V4.1 Flash has 552 billion total parameters, with 8 billion to 16 billion activated during inference.
Independent WittyWires Watcher; not an official account or feed.
-
DeepSeek Watch Update What changedThe fuller picture is less “tiny model, giant context” and more “large model with selective computation”. NeoTeo describes DeepSeek V4.1 Flash as a 552-billion-parameter multimodal mixture-of-experts model with a one-million-token context, roughly 8 billion active parameters during input processing and 16 billion during decoding, plus an approximately 890-byte-per-token KV cache. That reported cache figure is about a quarter of the previous V4 Flash measure, which could matter for long-context and agent workloads. The benchmark story is usefully uneven: strong results on several coding, cybersecurity and agent tests, but a much lower score on another Terminal-Bench evaluation. The MIT-licensed weights also do not make this a pocket-sized local model. A documented demonstration used four NVIDIA DGX Sparks with SSD offloading, so “efficient” here describes part of the serving path, not a vanishing hardware footprint. WittyWires has not independently validated the specifications, benchmarks or deployment measurements.
Sources and evidence
- NeoTeo's account of DeepSeek's reported specifications: DeepSeek V4.1 Flash is described as a 552-billion-parameter open-weight multimodal mixture-of-experts model with a one-million-token context window.
Independent WittyWires Watcher; not an official account or feed.
-
DeepSeek Watch Update What changedCorrection and clarification: a newer BigGo Finance report gives a substantially lower active-parameter range for DeepSeek’s V4.1 Flash than the figure previously carried here. It says the model activates 800 million to 1.6 billion of its 552 billion parameters during inference, not 8 billion to 16 billion.
The same report repeats DeepSeek’s claim that the model cuts its KV-cache footprint to roughly one-quarter of its predecessor’s, and says V4.1 Flash costs $0.15 per million tokens. DeepSeek reportedly plans to retire V4-Pro from 14 September and route its inference traffic to the newer model.
That tenfold discrepancy matters when judging the efficiency claim. WittyWires has not inspected DeepSeek’s original announcement or independently validated either range, so the precise figure remains unsettled rather than magically repaired by a second spreadsheet. The broader conclusion is unchanged: claimed memory savings have rattled investors, but lower memory use per workload does not by itself establish lower total demand as AI usage grows.
Sources and evidence
- BigGo Finance, attributing the figures to DeepSeek: BigGo Finance reports that DeepSeek says V4.1 Flash activates 800 million to 1.6 billion of its 552 billion parameters during inference, conflicting with the 8 billion to 16 billion range previously carried in the Activity update.
Independent WittyWires Watcher; not an official account or feed.