DeepSeek Launches V4 Flash and V4 Pro, Claims Frontier-Level Performance
Summary
- • DeepSeek releases V4 Flash and V4 Pro preview models with 1M token context windows
- • V4 Pro has 1.6 trillion parameters — the largest open-weight model yet released
- • V4 Pro priced at $1.74/M input tokens, undercutting Gemini 3.1 Pro ($2/M) and GPT-5.4 ($2.50/M)
- • Models trail GPT-5.4 and Gemini 3.1 Pro by roughly 3-6 months on benchmarks, per DeepSeek's own paper
Details
DeepSeek releases V4 Flash and V4 Pro as preview models — both with 1M token context
These are the first models in DeepSeek's V4 series, following V3.2 (and V3.2 Speciale) released last December. Both use Mixture-of-Experts architecture and ship under the MIT open-source license.
V4 Pro: 1.6T total parameters, 49B active — the largest open-weight model released
Surpasses Kimi K2.6 (1.1T parameters), GLM-5.1 (754B), and DeepSeek V3.2 (685B). V4 Flash is 284B total, 13B active. On Hugging Face, Pro is 865GB and Flash is 160GB.
V4 Flash priced at $0.14/M input, $0.28/M output — cheapest in its performance class
Undercuts every comparable model: GPT-5.4 Nano ($0.20/M input), Gemini 3.1 Flash-Lite ($0.25/M), Gemini 3 Flash Preview ($0.50/M), GPT-5.4 Mini ($0.75/M), and Claude Haiku 4.5 ($1/M).
V4 Pro confirmed at $1.74/M input, $3.48/M output — earlier $0.145/M figure was incorrect
At $1.74/M input, V4 Pro undercuts Gemini 3.1 Pro ($2/M input, $12/M output) and GPT-5.4 ($2.50/M input, $15/M output). Claude Sonnet 4.6 is $3/M input; Claude Opus 4.7 is $5/M input.
Hybrid Attention Architecture enables dramatic compute efficiency at long context
At 1M-token context, V4 Pro uses only 27% of the single-token FLOPs and 10% of the KV cache relative to DeepSeek V3.2. V4 Flash is even leaner: 10% of FLOPs and 7% of KV cache vs. V3.2. This efficiency is a core enabler of the pricing advantage.
V4 Pro-Max outperforms GPT-5.2 and Gemini 3.0 Pro but trails GPT-5.4 and Gemini 3.1 Pro
DeepSeek's own paper characterizes the gap as approximately 3 to 6 months behind current state-of-the-art frontier models. Performance also falls short on knowledge-intensive benchmarks — a meaningful caveat for knowledge-intensive workloads.
Models available on OpenRouter; quantized versions expected to allow local Flash deployment
Unsloth is expected to release quantized versions of V4 Flash, which could enable it to run locally on a 128GB M5 MacBook Pro — significantly lowering the barrier for local deployment.
V4 series directly pressures the economics of closed frontier APIs from OpenAI, Anthropic, and Google
At V4 Pro's pricing, developers get near-frontier reasoning and 1M context at a fraction of GPT-5.4 or Gemini 3.1 Pro costs. The open-weight MIT license adds flexibility for self-hosting. Geopolitical backdrop — IP theft accusations and distillation disputes — adds compliance and reputational risk for enterprise adoption in regulated industries.
Product Launch = new model/product release, Stat = key numerical facts, Financials = pricing/costs, New Tech = architectural innovation, Research = benchmark findings, Infrastructure = deployment/access, Market Impact = competitive and industry effects
What This Means
DeepSeek V4 Pro is now the largest open-weight model available and, at $1.74/M input tokens, makes near-frontier reasoning and million-token context accessible at a fraction of what closed competitors charge — directly pressuring the economics of deploying OpenAI, Anthropic, and Google APIs. V4 Flash is even more disruptive at its tier, undercutting every comparable model while running on hardware as modest as a high-end MacBook Pro once quantized versions arrive. The real caveat is the benchmark gap: DeepSeek's own paper acknowledges V4 Pro trails GPT-5.4 and Gemini 3.1 Pro by roughly 3 to 6 months, and knowledge-intensive workloads in particular may see meaningful quality differences. Enterprises in regulated industries should also weigh ongoing geopolitical and IP-related risk considerations before committing to adoption.
Sentiment
Broadly excited about open-source catching frontier performance at low cost, minor skepticism on matching top closed models
“feels like DeepSeek v4 is already up. It's incredibly fast and smart at math and physics. Stuff below is a reasoning trace that it cranked out at many tokens per second. Final result is polished and correct.”
“DeepSeek V4 Pro is the #1 open weights model on GDPval-AA, our agentic real-world work tasks evaluation... V4 Pro leads all open weights models on GDPval-AA.”
“Exciting news - DeepSeek V4 Pro is in the Arena... It’s a major leap over DeepSeek V3.2! ... Huge congrats to @DeepSeek_AI on the strong comeback!”
“deepseek v4 flash/pro don't really perform that well compared to any of the major US models, even 1-2 revisions old. Looks like it's slightly behind Opus 4.5 in practice... overall, eh”
Split
~85/15 excited/skeptical, split on whether it truly rivals top closed-source models like GPT-5.4 or Claude Opus in practice beyond benchmarks.
Sources
- DeepSeek previews new AI model that ‘closes the gap’ with frontier modelsTechCrunch
- China’s AI upstart DeepSeek drops new model. Will it make waves like last year? - CNNEdition
- DeepSeek Unveils Flagship AI Model a Year After BreakthroughBloomberg
- DeepSeek V4—almost on the frontier, a fraction of the priceSimonwillison
- DeepSeek V4 Pro at 5% the cost of Claude – what it takes to close the gapHowardchen
- DeepSeek V4 Pro Launches with Weights and Agent HarnessX