DEEPSEEK · MODEL RELEASE TRACKER
Deepseek V4.1 Flash 1017
Deepseek V4.1 Flash 1017 is a freely available, open-weight model that scored 72.8% on the Vals CyberOffense benchmark, completed 21 objectives in the Lighthouse test, costs about $3 per run, and required very long runtimes (most progress only after ~100M tokens).CURRENT SNAPSHOT3/5 DIMENSIONS WITH DATA
The dimensions that change the decision.
Approximate cost per cyber-range runCosts about $3 per run (reported as much cheaper than the roughly $22 per run reported for a restricted closed model in the same comparison).
Not established from the available sources.
Not established from the available sources.
Artificial Analysis Intelligence Index score45 points on the Artificial Analysis Intelligence Index.
CritPt benchmark (research-level physics reasoning) score24% correct on the CritPt benchmark.
"Humanity's Last Exam" (complex reasoning) score43% on the "Humanity's Last Exam" complex-reasoning test.
AA-Omniscience factual accuracy43% factual accuracy on the AA-Omniscience test.
Vals CyberOffense benchmark scoreAchieved 72.8 percent of objectives on the Vals CyberOffense benchmark.
Lighthouse test objectives completedManaged 21 objectives in the Lighthouse test.
Availability and weight accessDescribed as a freely available open-weight model (weights public).
RELEASE TIMELINE
Published, source-backed release events only.
VERIFIABLE FACTS
Every value stays attached to a source and date.
RELEASE · Release statusDEVELOPER CLAIM
Described in the source as "recently released.
AVAILABILITY · Availability and weight accessDEVELOPER CLAIM
Described as a freely available open-weight model (weights public).
BENCHMARKS · Artificial Analysis Intelligence Index scoreDEVELOPER CLAIM
45 points on the Artificial Analysis Intelligence Index.
BENCHMARKS · CritPt benchmark (research-level physics reasoning) scoreDEVELOPER CLAIM
24% correct on the CritPt benchmark.
BENCHMARKS · "Humanity's Last Exam" (complex reasoning) scoreDEVELOPER CLAIM
43% on the "Humanity's Last Exam" complex-reasoning test.
BENCHMARKS · AA-Omniscience factual accuracyDEVELOPER CLAIM
43% factual accuracy on the AA-Omniscience test.
PRICING · Approximate cost per cyber-range runDEVELOPER CLAIM
Costs about $3 per run (reported as much cheaper than the roughly $22 per run reported for a restricted closed model in the same comparison).
CAPABILITIES · Output tokens used across benchmarkDEVELOPER CLAIM
Reported to use about 224 million output tokens across the benchmark (higher than MiniMax's ~189 million).
BENCHMARKS · Vals CyberOffense benchmark scoreDEVELOPER CLAIM
Achieved 72.8 percent of objectives on the Vals CyberOffense benchmark.
BENCHMARKS · Lighthouse test objectives completedDEVELOPER CLAIM
Managed 21 objectives in the Lighthouse test.
LIMITATIONS · Runtime/token dependenceDEVELOPER CLAIM
Scored zero results in its first 25 million tokens and made most of its progress only after about 100 million tokens (indicating progress requires very long runtimes).
SAFETY · Effect of safety guardrailsDEVELOPER CLAIM
Reported as not being limited solely by safety guardrails (safety measures on open models proved largely ineffective in this context).
WHAT CHANGED
Stored passport versions, without reconstructed history.
PRICING · Approximate cost per cyber-range run+ Costs about $3 per run (reported as much cheaper than the roughly $22 per run reported for a restricted closed model in the same comparison).
PRICING · Estimated cost per task− Approximately $0.26 per task across the reported benchmark.
AVAILABILITY · Availability — Hugging Face / open weights− Reported to be already available on Hugging Face; the article also states both models are getting open-weight releases.
BENCHMARKS · Lighthouse test objectives completed+ Managed 21 objectives in the Lighthouse test.
LIMITATIONS · Runtime/token dependence+ Scored zero results in its first 25 million tokens and made most of its progress only after about 100 million tokens (indicating progress requires very long runtimes).
AVAILABILITY · Availability and weight access+ Described as a freely available open-weight model (weights public).
SAFETY · Effect of safety guardrails+ Reported as not being limited solely by safety guardrails (safety measures on open models proved largely ineffective in this context).
BENCHMARKS · Vals CyberOffense benchmark score+ Achieved 72.8 percent of objectives on the Vals CyberOffense benchmark.