Command Palette
Search for a command to run...

DeepSeek Previews Huawei-Chip AI Model as V4 Debuts

aiai-productsai-modeling 12 posts · 6 accounts

DeepSeek launched its V4 Flash and V4-Pro models and previewed a new AI model adapted to run on Huawei chips, pairing a product refresh with a move to support Huawei hardware.

DeepSeek’s API docs show V4-Pro is temporarily discounted to $0.435 per 1 million cache-miss input tokens and $0.87 per 1 million output tokens until May 5. A translated technical document linked with the rollout indicated MXFP4 support for Huawei’s 950DT chip, with training software planned within a week.

From the sources (12 posts)

@firstsquawk

DeepSeek previews new AI model adapted to run on Huawei chips

@poezhao0605

DeepSeek released V4 this week. The model is impressive. But the technical report is more interesting than the benchmarks. DeepSeek spends significant space telling hardware vendors what useful compute should look like. Specific compute-to

@teortaxestex

I checked the translation. This is an important document, our best window into what's actually going on between Huawei, DeepSeek and 950 series chips. One immediate thing: yes, MXFP4 support is for 950DT. They've got them. Training [softwa

@teortaxestex

Hey hey hey this is actually bigger than I thought. 950DT, not 950PR. Volume production of it was scheduled for *the end of 2026*. So DeepSeek really is one of the first customers of a relatively modern, general-purpose Chinese NPU. Physica

@teortaxestex

For the next 10 days, DeepSeek serves V4-Pro at a 75% discount, $0.43/$0.87 for 1M tokens (reminder that their costs are weird in $ because they're integers in ¥). This, I think, is close to its actual breakeven price given average context

@teortaxestex

An out-there hypothesis: V4 was supposed to use Engram, but in late 2025-early 2026 they were promised accelerated development of Ascend 950DT hardware, and redacted parts that were hard to adopt. Pro's post-training was rushed on Ascend,

@mweinbach

@zephyr_z9 And the model arch wasn’t really working for v4 pro they couldn’t even train the model more if they wanted to

@teortaxestex

Interesting that MathArena doesn't replicate DeepSeek's scores, in fact on Apex the score is 10% below the paper's. in comparison, Kimi K2.6's first party eval was 27.9%, which ofc is 4% above but a more plausible margin of error. Even FLAS

@teortaxestex

Of course, "expensive" is just a market reality. V4-Pro costs 8.28x more per output token vs Speciale, and 13.56x times more per experiment. So we can (loosely) infer it uses 63% more tokens to precisely triple the score. Its actual inferen

@scaling01

it's finally here (I'm such an idiot I forgot to post this when I first found DeepSeek-V4)

@lmsysorg

RT @0xishand: Thanks to an amazing partnership with @inferact and @lmsysorg /@radixark , Dynamo had day0 support for DeepSeek-V4 with featu…

@scaling01

DeepSeek is offering a 75% discount until May 5th

Preview built on a synthetic news corpus (16 weeks, Apr–Jul 2026). Impact calls are model reads, not price data.

About Archive