DEEPSEEK · MODEL RELEASE TRACKER
DeepSeek-V4-Flash
Reportedly, a forensic analysis by researcher Will DePue found statistical fingerprints in DeepSeek-V4-Flash's published model weights that are consistent with Nvidia's Blackwell-generation GPUs; DePue said he reproduced the result on real Blackwell chips and published verification code.CURRENT SNAPSHOT4/5 DIMENSIONS WITH DATA
The dimensions that change the decision.
PricingDescribed by DeepSeek as having 'highly cost‑effective API pricing' alongside smaller parameter size and faster response times.
Evaluated at million-token contextEvaluations and reported speedups reference DeepSeek-V4-Flash operating with 1,000,000-token contexts (1M-token sequences).
Not established from the available sources.
BioPhys-Bridge evidence-ID F1In preliminary BioPhys‑Bridge evaluations, DeepSeek‑V4‑Flash obtained the highest evidence‑ID F1 score of 0.360.
GameASG-Bench strict task successesOn GameASG‑Bench, with full tool access and larger nominal turn budgets DeepSeek‑V4‑Flash achieved 18 strict task successes per tested harness; only ten tasks succeeded under both harnesses.
IneqAgent benchmark success rateOn four evaluated benchmarks within the IneqAgent study, IneqAgent with DeepSeek-V4-Flash solved 175 of 186 problems (94.09%), achieving the highest overall success rate among the evaluated methods.
HyScale speedup on CSA layers (1M-token, 32 H800 GPUs)HyScale speeds up the CSA layers of DeepSeek-V4-Flash by 2.01× when evaluated at 1,000,000-token sequences on 32 H800 GPUs.
Forensic evidence of Nvidia Blackwell GPU usageA forensic analysis by Will DePue reported that DeepSeek-V4-Flash's published weights contain statistical patterns characteristic of Nvidia's Blackwell-generation GPUs; DePue said he reproduced the results on real Blackwell chips and published the verification code on GitHub, suggesting the model was trained using Blackwell hardware.
RELEASE TIMELINE
Published, source-backed release events only.
VERIFIABLE FACTS
Every value stays attached to a source and date.
OTHER · Model size (total / active parameters)DEVELOPER CLAIM
Reported as 284B total parameters and 13B active parameters.
CONTEXT WINDOW · Evaluated at million-token contextINDEPENDENTLY SUPPORTED
Evaluations and reported speedups reference DeepSeek-V4-Flash operating with 1,000,000-token contexts (1M-token sequences).
RELEASE · Reported release timeframe (grouping)DEVELOPER CLAIM
DeepSeek-V4-Flash was reported as one of six Chinese models released since April 2026 that carry the Blackwell signature, according to the same forensic analysis.
OTHER · Comparative reasoning and agent performanceDEVELOPER CLAIM
Reportedly 'reasoning capabilities closely approach V4‑Pro' and 'performs on par with V4‑Pro on simple Agent tasks' (per DeepSeek).
PRICING · PricingDEVELOPER CLAIM
Described by DeepSeek as having 'highly cost‑effective API pricing' alongside smaller parameter size and faster response times.
OTHER · Retirement and compatibility routingDEVELOPER CLAIM
DeepSeek states 'V4‑Flash & V4‑Flash‑Vision‑Exp are retired'; requests to deepseek‑v4‑flash and deepseek‑v4‑flash‑vision‑exp are temporarily routed to V4.1‑Flash for compatibility.
OTHER · GameASG‑Bench strict task successesINDEPENDENTLY SUPPORTED
On GameASG‑Bench, DeepSeek‑V4‑Flash achieved 18 strict task successes for each tested harness (with full tool access and larger nominal turn budgets), but only 10 tasks succeeded under both harnesses; strict task success rate was reported not monotonic in reasoning effort.
OTHER · BioPhys‑Bridge evidence‑ID F1INDEPENDENTLY SUPPORTED
Reported to obtain the highest evidence‑ID F1 score on BioPhys‑Bridge: 0.360.
CAPABILITIES · Reasoning and agentic performance relative to V4-ProDEVELOPER CLAIM
Reportedly, reasoning capabilities closely approach DeepSeek‑V4‑Pro; performs on par with V4‑Pro on simple agent tasks.
CAPABILITIES · Parameter countsDEVELOPER CLAIM
Reported as 284 billion total parameters and 13 billion active parameters.
BENCHMARKS · BioPhys-Bridge evidence-ID F1INDEPENDENTLY SUPPORTED
In preliminary BioPhys‑Bridge evaluations, DeepSeek‑V4‑Flash obtained the highest evidence‑ID F1 score of 0.360.
BENCHMARKS · GameASG-Bench strict task successesINDEPENDENTLY SUPPORTED
On GameASG‑Bench, with full tool access and larger nominal turn budgets DeepSeek‑V4‑Flash achieved 18 strict task successes per tested harness; only ten tasks succeeded under both harnesses.
AVAILABILITY · Forensic evidence of Nvidia Blackwell GPU usageDEVELOPER CLAIM
A forensic analysis by Will DePue reported that DeepSeek-V4-Flash's published weights contain statistical patterns characteristic of Nvidia's Blackwell-generation GPUs; DePue said he reproduced the results on real Blackwell chips and published the verification code on GitHub, suggesting the model was trained using Blackwell hardware.
BENCHMARKS · IneqAgent benchmark success rateINDEPENDENTLY SUPPORTED
On four evaluated benchmarks within the IneqAgent study, IneqAgent with DeepSeek-V4-Flash solved 175 of 186 problems (94.09%), achieving the highest overall success rate among the evaluated methods.
CAPABILITIES · Performance at low inference costINDEPENDENTLY SUPPORTED
In the study, across four backbone models including DeepSeek-V4-Flash, IneqAgent using DeepSeek-V4-Flash achieved higher success rates at low LLM inference cost thresholds than a Direct Agent using the same backbone, and exhibited lower mean inference costs over each method's successful runs.
BENCHMARKS · HyScale speedup on CSA layers (1M-token, 32 H800 GPUs)INDEPENDENTLY SUPPORTED
HyScale speeds up the CSA layers of DeepSeek-V4-Flash by 2.01× when evaluated at 1,000,000-token sequences on 32 H800 GPUs.
CAPABILITIES · Contains CSA layersINDEPENDENTLY SUPPORTED
The model is described as having CSA layers (a learned-sparse-attention component targeted by HyScale optimizations).
WHAT CHANGED
Stored passport versions, without reconstructed history.
AVAILABILITY · Forensic evidence of Nvidia Blackwell GPU usage+ A forensic analysis by Will DePue reported that DeepSeek-V4-Flash's published weights contain statistical patterns characteristic of Nvidia's Blackwell-generation GPUs; DePue said he reproduced the results on real Blackwell chips and published the verification code on GitHub, suggesting the model was trained using Blackwell hardware.
RELEASE · Release & availability− Declared as part of the DeepSeek‑V4 Preview, described as live and open‑sourced, with the API updated and available.
RELEASE · Reported release timeframe (grouping)+ DeepSeek-V4-Flash was reported as one of six Chinese models released since April 2026 that carry the Blackwell signature, according to the same forensic analysis.
AVAILABILITY · Availability / routing− DeepSeek‑V4‑Flash and V4‑Flash‑Vision‑Exp are retired; for compatibility, requests to deepseek‑v4‑flash are temporarily routed to DeepSeek‑V4.1‑Flash.
CONTEXT WINDOW · Default context length− Default context length is 1,000,000 tokens (1M context is stated as the default across official DeepSeek services).
CAPABILITIES · Contains CSA layers+ The model is described as having CSA layers (a learned-sparse-attention component targeted by HyScale optimizations).
CONTEXT WINDOW · Evaluated at million-token context+ Evaluations and reported speedups reference DeepSeek-V4-Flash operating with 1,000,000-token contexts (1M-token sequences).
BENCHMARKS · HyScale speedup on CSA layers (1M-token, 32 H800 GPUs)+ HyScale speeds up the CSA layers of DeepSeek-V4-Flash by 2.01× when evaluated at 1,000,000-token sequences on 32 H800 GPUs.
BENCHMARKS · IneqAgent benchmark success rate+ On four evaluated benchmarks within the IneqAgent study, IneqAgent with DeepSeek-V4-Flash solved 175 of 186 problems (94.09%), achieving the highest overall success rate among the evaluated methods.
CAPABILITIES · Performance at low inference cost+ In the study, across four backbone models including DeepSeek-V4-Flash, IneqAgent using DeepSeek-V4-Flash achieved higher success rates at low LLM inference cost thresholds than a Direct Agent using the same backbone, and exhibited lower mean inference costs over each method's successful runs.
BENCHMARKS · BioPhys-Bridge evidence-ID F1+ In preliminary BioPhys‑Bridge evaluations, DeepSeek‑V4‑Flash obtained the highest evidence‑ID F1 score of 0.360.
OTHER · BioPhys‑Bridge evidence‑ID F1Reported to obtain the highest evidence‑ID F1 score on BioPhys‑Bridge: 0.360.→Reported to obtain the highest evidence‑ID F1 score on BioPhys‑Bridge: 0.360.
OTHER · Comparative reasoning and agent performanceReportedly 'reasoning capabilities closely approach V4‑Pro' and 'performs on par with V4‑Pro on simple Agent tasks' (per DeepSeek).→Reportedly 'reasoning capabilities closely approach V4‑Pro' and 'performs on par with V4‑Pro on simple Agent tasks' (per DeepSeek).
CONTEXT WINDOW · Default context lengthReported support for a 1,000,000‑token (1M) context as the default across DeepSeek V4 services; model supports dual modes (Thinking / Non‑Thinking).→Default context length is 1,000,000 tokens (1M context is stated as the default across official DeepSeek services).
PRICING · PricingDescribed by DeepSeek as having 'highly cost‑effective API pricing' alongside smaller parameter size and faster response times.→Described by DeepSeek as having 'highly cost‑effective API pricing' alongside smaller parameter size and faster response times.
BENCHMARKS · GameASG-Bench strict task successes+ On GameASG‑Bench, with full tool access and larger nominal turn budgets DeepSeek‑V4‑Flash achieved 18 strict task successes per tested harness; only ten tasks succeeded under both harnesses.
OTHER · GameASG‑Bench strict task successesOn GameASG‑Bench, DeepSeek‑V4‑Flash achieved 18 strict task successes for each tested harness (with full tool access and larger nominal turn budgets), but only 10 tasks succeeded under both harnesses; strict task success rate was reported not monotonic in reasoning effort.→On GameASG‑Bench, DeepSeek‑V4‑Flash achieved 18 strict task successes for each tested harness (with full tool access and larger nominal turn budgets), but only 10 tasks succeeded under both harnesses; strict task success rate was reported not monotonic in reasoning effort.
CAPABILITIES · Parameter counts+ Reported as 284 billion total parameters and 13 billion active parameters.
RELEASE · Release & availabilityDeclared as part of the DeepSeek‑V4 Preview, described as live and open‑sourced, with the API updated and available.→Declared as part of the DeepSeek‑V4 Preview, described as live and open‑sourced, with the API updated and available.
OTHER · Model size (total / active parameters)Reported as 284B total parameters and 13B active parameters.→Reported as 284B total parameters and 13B active parameters.
CAPABILITIES · Reasoning and agentic performance relative to V4-Pro+ Reportedly, reasoning capabilities closely approach DeepSeek‑V4‑Pro; performs on par with V4‑Pro on simple agent tasks.
OTHER · Retirement and compatibility routingDeepSeek states 'V4‑Flash & V4‑Flash‑Vision‑Exp are retired'; requests to deepseek‑v4‑flash and deepseek‑v4‑flash‑vision‑exp are temporarily routed to V4.1‑Flash for compatibility.→DeepSeek states 'V4‑Flash & V4‑Flash‑Vision‑Exp are retired'; requests to deepseek‑v4‑flash and deepseek‑v4‑flash‑vision‑exp are temporarily routed to V4.1‑Flash for compatibility.
AVAILABILITY · Availability / routing+ DeepSeek‑V4‑Flash and V4‑Flash‑Vision‑Exp are retired; for compatibility, requests to deepseek‑v4‑flash are temporarily routed to DeepSeek‑V4.1‑Flash.