In factorial design experiments we inoculated fresh salmon fillets with a spore pool of 13 nonproteolytic strains of Clostridium botulinum type B, E, and F at 6 levels (10−1 to 104/50 g of fillet), and incubated at 1, 4, 8, 12 and 30°C under modified atmospheres (MA) of vacuum, 100% CO2 and 70% CO2 + 30% air for up to 60 d. The earliest time we detected toxin in the fillets at 30, 12 and 8°C, irrespective of MA, was after 1, 3–9 and 6–12 d of storage and required 100–103, 101–103, 101–102 spores/fillet. The probability (P) of toxin production was significantly (P<0.05) affected by temperature (T), MA storage time (ST), MA × T, MA × ST and T × ST. Only type B toxin was detected in the toxic fillets. No toxin was detected in fillets stored at 4°C for up to 60 d. Toxin detection coincided with spoilage at 30°C, but preceded spoilage at 8 and 12°C, and followed spoilage at 4°C. Using linear and logistic regression analysis, best fit equations were derived relating the length of the lag phase and P of toxin production to T, ST, MA and spore inoculum level.