python 3.8, 161+1+287+269+6985=7702

from zlib import decompress as D
from itertools import product as P
from functools import reduce as R
x=287
d=open('f','rb').read()
print(eval(D(d[:x]))(d[x:]))

where D(d[:x]) is

(lambda f,
rest_len=[265,7,8,24,50,10,6,134,14,24,229],
genes_len=[7097,1274,276,76,223,62,122,122,420,39],
cut=lambda s,x:R(
    lambda acc,n:[acc[0][n:],acc[1]+[acc[0][:n]]],
    x,[s,[]])[1],
cut_out=lambda x,n:x[:n]+x[n+1:],
C=[''.join(i) for i in P('AGCU',repeat=3)]:
cut_out(''.join(i+''.join(j)
    for i,j in zip(
        cut(D(f[:269]).decode('utf-8'),rest_len),
        cut([C[i-32] for i in D(f[269:])],genes_len)+[[]])
    ),13468)
)

and the file f is here (produced with decode.py).

Posted here only to show grouping nucleotides approach as the decoder overhead is very large.
So let's start with the explanations. As can be seen form MN908947, the virus mRNA contains "head" (265 nucleotides), "tail" (229 nucleotides) and translated parts with some non-translated parts between them. So, we have 10 genes, named orf1ab, S, ORF3a, E, M, ORF6, ORF7a, ORF8, N, ORF10 in the link above.
It comes out we could use codon usage bias as the picture of frequency codon (=nucleotide triples) usage is very impressive, e.g. for orf1ab:

or for S (looks very similar):

and for N (looks different):
More precisely, here's the table of normed covariance of the codon usage distributions:

       orf1ab S      ORF3a  E      M      ORF6   ORF7a  ORF8   N      ORF10 
orf1ab 1.0000 0.9243 0.7458 0.4237 0.4813 0.5143 0.6391 0.6897 0.5035 0.3201
S      0.9243 1.0000 0.7077 0.3834 0.4596 0.5266 0.6513 0.6451 0.5371 0.3519
ORF3a  0.7458 0.7077 1.0000 0.4683 0.5483 0.4119 0.5713 0.5957 0.3464 0.2862
E      0.4237 0.3834 0.4683 1.0000 0.2966 0.2057 0.4727 0.2994 0.0318 0.2621
M      0.4813 0.4596 0.5483 0.2966 1.0000 0.2495 0.3799 0.3432 0.2975 0.1776
ORF6   0.5143 0.5266 0.4119 0.2057 0.2495 1.0000 0.3737 0.3459 0.1980 0.3125
ORF7a  0.6391 0.6513 0.5713 0.4727 0.3799 0.3737 1.0000 0.5232 0.3223 0.2362
ORF8   0.6897 0.6451 0.5957 0.2994 0.3432 0.3459 0.5232 1.0000 0.2049 0.0902
N      0.5035 0.5371 0.3464 0.0318 0.2975 0.1980 0.3223 0.2049 1.0000 -0.0203
ORF10  0.3201 0.3519 0.2862 0.2621 0.1776 0.3125 0.2362 0.0902 -0.0203 1.0000

So we want to compress these genes. Here's the comparison of compression approaches:

         arithm  .lzma    .xz   zlib  a.ovh    len
0:orf1ab 4925.7   5261   5308   5083   86.5   7097
1:S       879.5   1047   1092    940   59.1   1274
2:ORF3a   191.4    276    324    237   37.4    276
3:E        47.8    100    132     84   18.6     76
4:M       155.4    232    280    200   35.1    223
5:ORF6     36.9     86    120     70   15.4     62
6:ORF7a    81.2    143    184    124   25.2    122
7:ORF8     80.2    144    184    122   24.2    122
8:N       289.9    393    440    338   44.1    420
9:ORF10    22.5     64     96     47   11.6     39

where arithm is arithmetical encoding and a.ovh is frequences compressed with fibonacci (+1 ending bit for each) -- the arithmetical encoding overhead.
We see the comparison makes zlib a perfect candidate, as even if we compress all with arithmetical encoding, we would have left something like 193 bytes for writing the decoding overhead (versus zlib), which is highly unlikely to fit in there.